People Detection Optimization from a Partial Gait Cycle

1. Introduction

Pattern recognition approaches have gained a lot of success in the field of visual recognition. The mathematical foundations of these approaches, particularly scale-invariant feature extraction and probabilistic modeling, draw from principles in statistical physics and fractal geometry. Examples include face detection, vehicle license plate recognition, fingerprint recognition, and others. Each of these approaches uses machine learning methods to build a diagnostic from a large dataset. The detector then provides a sample of all possible images as input to the target object. The results and experience show that these systems have acceptable use and efficiency for various tasks in industry, agriculture, military, etc. One of the similar parts in this regard is the pedestrians detection from images. Most of the methods presented in this area have difficulties in pedestrians detection due to the abundance of objects in the images. In addition, it is difficult to detect pedestrians because of poor quality urban cameras. In general, because of their low quality, city cameras typically contain between 100 and 200 pixels of target for pedestrians [2, 1]. Because of this, pedestrian detection is not possible with classic machine vision techniques. Therefore, this study attempts to present an automated method for processing images simultaneously using video frames based on the principles of machine vision and updated machine learning.

Detection algorithms should be fast, efficient, and searchable among the input data in any dimension [1]. Input data must also be large enough to detect detection rates. The way forward research is to use machine learning and advanced image processing techniques that are based on machine vision techniques. In fact, deep learning techniques which use VGG-NET based on CRF method is intended to be used to train the entire set of images as input data. The way to do this is to first improve the set of input images. Image enhancement is done in terms of brightness, resizing to a certain size, eliminating probable noise and others. Then, by considering image segmentation, features can be extracted. For segmentation, a special method called fractal is used, which has three important properties, including self-similarity, lack of dimension and unique. These features follow specific operations with respect to the edges of the image to identify the boundary points trained by pedestrians.

The combination of fractal theory and deep learning models, therefore, enables a new perspective for partial gait recognition where scale-invariance becomes a concern. Fractal models have self-similarity and non-integer dimensional characteristics, making it an intuitive model to express the hierarchical structure of human motion characteristic at multiple time scales. The main contributions of this research are: (1) Implementation of VGG-Net-CRF model reengineered to include fractal segmentation, which improves gait image feature extraction, (2) Formulation of the fractal based dimensionality reduction approach for gait recognition process, (3) Both benchmark performance up to 97.40% and real-time accuracy testing while operating under degraded surveillance conditions that represents low quality scenario, and (4) Rigorous statistical validation using multiple sets of benchmark data with confidence interval for testing and classification.

2. Literature Review

There has not been much research into the pedestrian detection of surveillance camera images or videos using machine learning techniques to date. In [1], deep convolutional neural network is discussed which performs better than other similar methods. In another approach presented in [2], pedestrians are distinguished by their movement and appearance. Input datasets are trained using the Adaboost Neural Network method and the pedestrian detection approach is based on the Viola Jones algorithm, which has been used extensively in detection methods. In fact, the authors of this research have developed the Viola-Jones algorithm, which has been applied in various fields. In another approach presented in [3], they have identified the body states of people while walking as pedestrian detection. The use of spatial layers is taken into account considering the pedestrian distance to the linear boundary of the road. Detecting the people take to determine the motion of an object in a scene considered a feature. Using the method of Canny edge detection with Hough transform algorithm is the method presented in this study.

The use of motion picture history as a motion-based filtering of people in the scene has also been used to detect pedestrians from infrared images [4]. Another research used convolution-based learning methods that works to classify with convolutional neural network and deep learning after spatial domain feature extraction [5]. The use of multi-sparse descriptors in imagery with the construction of a glossary of visual features has also been proposed [6]. In another approach, the method of learning the hybrid mask model is used in the framework of multi-sample training using the support vector machine classification method [7]. In [8], pedestrians were also distinguished from infrared images using the Sparse method and the HOG algorithm, the results of which are striking compared to previous methods.

Accurate, robust, and real-time pedestrian detection using deep learning provided in [9]. In addition to identifying pedestrians' movement, the study found that their movement position was not addressed in previous articles. The use of a convolutional-based neural network based on the descriptor of added channel features considered as the main solution. The simulation results represented the high accuracy of pedestrian detection and the observational results showed a somewhat appropriate detection in cameras and robots. The use of a robust multispectral mode to detect pedestrians from video in [10] was also based on deep learning and the use of a neural network convolution technique called the R-CNN method. The method was fast, efficient and robust. In general, Table 1 summarizes the advantages and disadvantages of previous methods.

Recent developments of gait-based recognition include Graph Convolutional Networks (GCN) that consider human pose as graph structures [18], attention-guided networks for occluded pedestrian detection [16] and FractalNet architectures for scale-invariant feature learning [19]. But none of these methods synthesizes fractal geometry with CRF optimization for partial gait cycle analysis.

Of course, there are other studies that are similar to the state-of-the-arts and novel methods that have been done, but have eliminated the previous weaknesses, but the advantages of the previous methods cannot be seen in all of them. Therefore, presenting and innovating a method to cover the weakness of previous methods and having their advantages is an important issue that the proposed approach of this research will do. For example, in [11] which used Asymmetric Multi-stage CNNs in small dimensions used which can detect pedestrians. This method named AMS-Net, the method uses data sets from INRIA, ETH, KITTI and CityPerson. In [12] a framework for recognizing the foreground of images and videos with the aim of detecting and tracking pedestrians from moving cameras in real time. The data used is CVC-14, and the overall approach is based on a feature extraction and classification mechanism with a Support Vector Machine (SVM) as a decision management system using two descriptors means HOG and LBP.

In [13], semantic pedestrian detection performed from a crowded population. The approach of this research is to identify the human heads area and then other parts, so this method is called Head-Body Alignment Net (HBAN) and works on CityPerson data. In [14], the convolutional neural network used for pedestrian detection in a multi-spectral manner on the CityPerson data set. Presentation of a new method of convolutional neural network — Region-based Fully Convolutional Networks or R-FCN based on a Network in Network or NIN — is considered in this research. In [15], the deep Parallel Feature Transfer Network or PftNet proposed and tested on INRIA, ETH and UMP2018 data. The use of Attention Guided Neural Network or AGNN in [16] is considered to be based on CityPerson data set. Also in [17], Fully Convolutional Neural Network (FCN) in deep learning model proposed for pedestrian identification from partial gait cycle which used CASIA-B as input dataset. This model obtained 94.10% accuracy with 0.21 seconds runtime.

3. Methodology

The VGG-Net architecture serves as our backbone CNN for feature extraction and pedestrian detection. VGG-Net consists of sequential convolutional layers with 3×3 filters, followed by max pooling layers with 2×2 kernels. We employ standard deep learning training procedures using gradient descent optimization with backpropagation to update all network weights. Our VGG-Net-CRF framework enhances the basic VGG architecture by integrating Conditional Random Field (CRF) optimization for spatial coherence modeling. The VGG backbone generates per-pixel classification scores that serve as unary potentials for the CRF. During training, we optimize the entire network end-to-end using standard SGD/Adam optimizers with learning rates adapted through scheduled decay. The network architecture consists of:

1. Input layer: 256×256×3 (RGB) or 256×256×1 (IR) images
2. VGG backbone: Sequential conv-pool blocks with filter depths [64, 128, 256, 512, 512]
3. CRF layer: Fully-connected pairwise potentials for spatial consistency
4. Output layer: Per-pixel pedestrian classification scores

All convolutional weights are initialized using Xavier initialization and updated through backpropagation during training. We use standard deep learning practices including data augmentation, dropout regularization, and batch normalization. The main reasons why to use a VGG-Net-CRF in this one instead of other intelligent methods that have used in a number in classification and feature extraction step as a Convolutional Neural Network and Recursive Neural Network — these two are deep learning techniques — and also conventional clustering techniques, namely Support Vector Machine and Naïve Bayesian, that are cited in Table 2.

VGG-Net follows standard deep learning architectures with hierarchical feature extraction through convolutional layers, similar to other CNN-based methods but enhanced with CRF spatial modeling. The VGG-Net algorithm can employ nonlinear activation functions such as sigmoid or sinusoidal or non-derivative activation functions as well as using a linear function to activate cells or neurons in the hidden layer because of its high flexibility. By default, VGG-Net has an equation in general mode such as equation (1).

where \(F^{l}\) represents feature maps at layer \(l\), \(W^{l}\) are learnable convolutional weights, \(b^{l}\) are bias terms, \(\sigma\) is the activation function (ReLU), and \(*\) denotes the convolution operation. The CNN forward pass processes input through sequential layers as follows. Each convolutional layer applies learned filters to extract hierarchical features, followed by non-linear activation and optional pooling for spatial down-sampling. The VGG-Net backbone architecture consists of:

• Convolutional layers: 3×3 kernels with ReLU activation
• Max pooling layers: 2×2 kernels with stride 2
• Feature map channels: [64, 128, 256, 512, 512] across blocks
• Batch normalization after each convolutional layer
• Dropout (p = 0.5) in fully connected layers

All network parameters are optimized using standard backpropagation with the Adam optimizer (learning rate = 1e-4, β₁ = 0.9, β₂ = 0.999). The network is trained end-to-end using gradient descent on the composite loss function.

For calculating the fitness function of the proposed CRF-VGG-Net based on the fractal model in this article, the dataset is considered as \(S \times D\), where \(S\) is a sample from each image and \(D\) is the sample's distance (features), which will have \(K\) segmented parts, and it is calculated as equation (2) [31].

In this equation, \(L\) is the distance (features) metric, taken as the Euclidean distance between any segmented parts, which is defined based on two features — brightness and edges. Finally, the flowchart of the proposed model is as Figure 1.

None of the prior methods synthesizes fractal geometry with CRF optimization for partial gait cycle analysis — this work connects classical fractal geometry with modern structured deep learning for surveillance-quality pedestrian analytics.

3.1 Data Preprocessing

A systematic pre-processing pipeline is used in our proposed method for normalising the input video sequences and to ease the feature extraction. The pre-processing framework consists of four important steps in which we improve the quality and coherency of the input data which is crucial for reliable pedestrian detection. Videos are first decomposed into frames at a fixed sampling rate of 30 frames per second to provide sufficient temporal resolution for capturing the complete attributes of the gait cycle, while keeping such a condition computationally feasible for real-time systems. Noise reduction is the second preprocessing step which is realized by a Gaussian filter with empirically optimized parameters. The filtering can be expressed mathematically as:

where \(f(x,y)\) is the original frame intensity and \(G(x,y,\sigma)\) is the Gaussian kernel which has a standard deviation \(\sigma = 1.2\), and \(\otimes\) is the convolution operator. This filtering technique efficiently suppresses sensor noise and compression artifacts due to the surveillance video system, and the important gait pattern features are retained. Illumination normalization aims to handle the problem of realistic surveillance environments with challenging backgrounds due to illumination changes. Histogram equalization is performed with the following transformation:

where \(L\) is the total number of discrete gray levels, \(h(j)\) is the frequency of the histogram at intensity level \(j\), and \(N\) is the total number of pixels. This operation is a way to achieve a similar brightness distribution in different kinds of conditions. The last preprocessing step is the spatial normalization, which performs image rescaling to a 256×256 pixel resolution by means of bilinear interpolation, in order to ensure common size of input frames under different camera systems.

3.2 Fractal Feature Extraction

The fractal model part takes advantage of the natural self-similarity property of human locomotive patterns over a range of temporal and spatial scales. This mathematical model exploits three basic fractal properties that characterize the geometric complexity and scale-invariance of gait dynamics. The fractal dimension measures the extent to which gait contours fill space over small scales, and represents a scale- and orientation-invariant measure of pattern complexity. The box-counting fractal dimension is calculated using the box-counting method, and the theoretical definition for this method is:

where \(N(\varepsilon)\) represents the minimum number of boxes with side length \(\varepsilon\) required to completely cover the extracted gait silhouette. For computational implementation, the practical approximation is utilized:

where \(\varepsilon_1\) and \(\varepsilon_2\) represent distinct box sizes within the optimal scaling range. Self-similarity quantification captures the invariance properties of gait patterns under geometric transformations. The self-similarity coefficient is computed as:

where \(I(x,y)\) is the intensity of a pixel at spatial coordinates \((x,y)\), and \(k\) is a scaling factor from the set \(\{0.5, 0.75, 1.25, 1.5\}\), and \(n\) is the number of total pixel pairs. This is a normalized measure of pattern consistency across scales, with values tending toward unity indicating high self-similarity. The non-dimensionality index measures the geometric complexity of gait boundary features by the formula:

where \(P\) denotes the perimeter length and \(A\) represents the enclosed area of the gait region. This dimensionless parameter captures the deviation from ideal circular geometry, providing insight into the irregular nature of human silhouette boundaries during locomotion.

3.3 VGG-Net-CRF Architecture

The implementation merges fractal-based feature representations in the context of a modified VGG-Net improved by Conditional Random Field optimization. This hybrid method combines deep convolutional networks for feature extraction together with the spatial coherence modeling of probabilistic graphical models. The network is organised with a concatenation of consecutive convolutional blocks with increasingly large filter depth (64, 128, 256, 512, 512) and a standalone fractal integrative layer to work with the computed geometric features. The Conditional Random Field (CRF) based optimization introduces fractal regularization with an augmented energy function formulated as:

The unary potential \(\psi_{u}(y_i) = -\log P(y_i \mid I)\) captures the classification confidence from the VGG-Net backbone, and \(\psi_{p}(y_i, y_j)\) enforces spatial consistency between neighbouring pixel classifications. The new fractal regularization term \(\psi_{f}(y_i, F_i)\) guarantees geometrical coherence by taking advantage of fractal feature constraints. The fractal regularizer is expressed as:

where \(\alpha\), \(\beta\) and \(\gamma\) are trainable weighting parameters, and \(D_{\text{ref}}\), \(S_{\text{ref}}\), \(NDI_{\text{ref}}\) are reference values calculated from the statistics of the training data. This penalty term punishes departures from the expected fractal behavior and accounts for classification results that are geometrically consistent with normal human gait. The pairwise potential is the enriched form that includes fractal-aware spatial relations:

where \(\mu(y_i, y_j) = [\,y_i \neq y_j\,]\) is the label compatibility function that penalizes different labels for similar pixels. The appearance kernel \(k^{(1)}\) encourages pixels with similar color and spatial proximity to have the same label, while the smoothness kernel \(k^{(2)}\) promotes spatial coherence regardless of appearance. The fractal kernel \(k_{\text{fractal}}\) incorporates geometric similarity based on fractal features. The fractal kernel is defined as:

promoting spatial coherence based on geometric feature similarity.

3.4 Training Optimization

The optimization framework employs a composite loss function that balances classification accuracy with geometric consistency constraints. The total loss function is expressed as:

where \(L_{\text{ce}} = -\sum_{i} y_i \log(\hat{y}_i)\) represents the standard cross-entropy loss for classification tasks. The fractal consistency term:

penalizes deviations between predicted and target fractal features, while \(L_{\text{crf}}\) represents the CRF energy minimization objective and:

The training method uses a multi-stage optimization scheme consisting of transfer learning with initial setting of VGG-Net weights pre-trained on ImageNet. Fractal feature extraction parameters are tuned by using gradient descent methods in the first place, then end-to-end joint optimized with the Adam optimizer and learning rate set to 1×10⁻⁴, momentum term values β₁ = 0.9, β₂ = 0.999. CRF parameters are trained with mean-field approximation, which is iterated until convergence, leading to fast inference at both training and testing. Hyperparameter tuning is performed using systematic grid search and cross-validation mechanisms. Fractal regularization weights (Table 4) are fine-tuned via empirical performance analysis: α = 0.1, β = 0.2, γ = 0.15, and CRF spatial weighting parameters have been set to λ₁ = 0.5, λ₂ = 0.3, λ₃ = 0.01. Training is continued for up to 100 epochs using early stopping where the training stops if 10 epochs do not make any improvement in validation. It is limited to 16 samples as the computation of fractal features is computationally expensive. Data augmentation includes random horizontal flipping (p = 0.5), random rotation (±10°), Gaussian noise (σ = 0.02), and random brightness adjustment (±0.2). The learning rate follows a cosine annealing schedule with warm restart every 30 epochs. Batch size is set to 16 due to memory constraints of fractal feature computation.

3.5 Real-time Implementation

Computational efficiency requirements for real-time pedestrian detection necessitate several optimization strategies implemented throughout the processing pipeline. Fractal feature computation utilizes GPU-accelerated parallel algorithms to minimize processing latency while maintaining mathematical precision. The CRF inference employs mean-field approximation with a fixed iteration count of 5, providing an optimal balance between computational speed and convergence accuracy. Temporal consistency is enhanced through a feature caching mechanism that stores previously computed fractal characteristics for consecutive frames, exploiting the temporal correlation inherent in video sequences. Model compression techniques include structured pruning that removes 15% of network connections with minimal impact on classification accuracy, and 8-bit quantization of feature maps to reduce memory bandwidth requirements. Memory management strategies implement a sliding window approach with a buffer capacity of 10 frames, enabling continuous video processing while maintaining bounded memory consumption. Gradient checkpointing during training phases reduces the memory footprint by selectively storing intermediate activations, allowing for larger effective batch sizes within hardware constraints. These optimizations collectively enable real-time processing capabilities with average inference times of 0.18 seconds per frame while maintaining the mathematical rigor of the fractal-enhanced CRF framework.

4. Experimental Design and Implementation

The performance of the proposed method is demonstrated on four different datasets to make the performance assessment more comprehensive and verify the generalizability. The CASIA-B dataset contains 124 subjects taken at 11 viewing angles and three walking types (normal, slow walking, fast walking) and provides 13,640 gait samples which are enough for the robust training and testing. The CASIA-C dataset includes 153 subjects with varying walking speeds and carrying conditions, 4,590 thermal sequences, and collected using infrared video sequences in different illumination conditions.

The CityPerson dataset, depicting real scenarios with 5,000 images of 35,000 annotated pedestrians in crowded streets, would be important to validate the deployment readiness. The INRIA Person Dataset is a standard benchmark consisting of 1,805 64×128 pixel pedestrian samples and 1,218 negative samples and allows a direct comparison to other methods. Dataset partition is based on a stratified 70-20-10 split among training, validation, and testing, maintaining balance in terms of subjects and environmental condition. Cross-validation uses a 5-fold approach with subject-independent splits to avoid identity bias and to evaluate generalization performance robustly.

4.1 Experimental Setup

Performance comparison follows standard procedures, and the same pre-processing is used for all the considered datasets. Ground truth annotations are controlled via manual inspection and inter-annotator agreement, with Cohen's kappa coefficients >0.85 for all datasets. Experiments are conducted on a high performance computing machine equipped with an Intel Core i7-12700K processor at 3.6 GHz (level-8 cache of 24 MB, 12 cores, 20 threads), 32 GB DDR4-3200 memory and an NVIDIA RTX 3080 Ti graphics card with 12 GB GDDR6X memory. Storage infrastructure is a Samsung 980 PRO NVMe 2 TB SSD with fast data access during both training and inference. Software is implemented on the PyTorch 1.12.1 framework with CUDA 11.7 support for GPU acceleration, on Ubuntu 20.04 LTS. Code was developed and tested using Python 3.9.7 and important libraries: NumPy 1.21.2, OpenCV 4.5.3, scikit-learn 1.0.2, and Matplotlib 3.4.3 for data processing and visualization. Fractal computation is based on custom CUDA kernels optimized for parallel operations of box counting, and CRF inference is performed using the DenseCRF library with mean-field approximation optimizations.

The reproducibility is guaranteed by setting fixed random seeds (seed = 42) for all random number generators, deterministic CUDA operations coding, and thorough version control using Git along with detailed commit logs. Training settings as well as hyperparameters are saved in YAML configuration files, making replication of exactly the same experiment trivial. Full code implementation along with documentation and code snippets for usage is available. Training/validation/test splits are fixed across all experiments: 70%/20%/10% for detection datasets, and subject-independent splits for gait datasets. We report results with 95% confidence intervals computed via bootstrap sampling (n = 1000 iterations). Statistical significance is assessed using paired t-tests with Bonferroni correction for multiple comparisons.

4.2 Quantitative Evaluation

Performance is evaluated using multiple criteria to capture various aspects such as classification accuracy, temporal coherence, and efficiency. For pedestrian detection, we report Average Precision (AP) at IoU thresholds 0.5 and 0.5:0.95, and Log-Average Miss Rate (LAMR) following Caltech evaluation protocols. For tracking evaluation, we compute Multiple Object Tracking Accuracy (MOTA), ID F1-score (IDF1), and Higher Order Tracking Accuracy (HOTA) metrics. Sensitivity analysis is calculated with different rates of occlusion (25%, 50%, 75%) by systematically masking regions, and noise resilience evaluation is performed across levels of variance of Gaussian noise (σ = 0.1–1.0). Among the temporal consistency metrics are tracking stability (i.e., IoU persistence across adjacent frames) and trajectory smoothness (i.e., acceleration variance). Performance evaluation pertains to inference time in FPS, memory usage during peak performance period, and energy efficiency measured on mobile devices.

Statistical significance testing is performed using one-way ANOVA with Bonferroni correction to allow for multiple comparisons, with family-wise type-1 error controlled at α = 0.05. Effect sizes are computed as Cohen's d for pair-wise comparison and eta-squared (η²) for overall model contributions. For all performance evaluations, the robust confidence interval estimation is stable when considering 1000 iterations of bootstrapping, as the performance is estimated over the uncertain data across datasets (i.e., dataset relationship). The inter-method assessment tests are structured as paired t-tests for dependent samples comparing identical test sets — this allows for proper statistical inference in terms of performance differences. Simulation results are presented in Figures 2 to 4 at the time of detection and tracking of pedestrians.

It is observed that a frame-by-frame and real-time tracking is being performed to detect and track pedestrians. It is clear that these outputs represent frames of 6, 38 and 60 frames. At the top-left side, the main frame and, at the bottom-left side, the frames that have been pre-processed. In the upper-right side, training and testing operations occur with the VGG-Net-CRF method and individuals are tracked in red. But their mobility causes errors in these red spots to remain; to prevent this red color redundancy for tracking, the fractal model to extract optimal features in each frame and adapt to previous frames and reduce the dimensionality of these features starts to work, and in some frames also shows a crowd of people walking, drawing green or red lines around that area. Other inputs for partial gait cycle pedestrian detection are shown in Figures 5 to 10. In terms of evaluation, the Precision-Recall curve and the Average Precision rate are shown in Figure 11 on the CityPersons and KAIST datasets.

Performance evaluation demonstrates strong detection and tracking performance with Average Precision scores consistently above 0.8 across different datasets. Similarly, other evaluation criteria are in Table 3.

Table 4 shows this comparison in terms of evaluation criteria to previous approaches to guarantee the results of the proposed method and display improvement under the same data usage conditions.

Table 5 shows the contribution of each component in our approach. The baseline VGG network achieves 80.4% AP@0.5 and 11.5% LAMR on the CityPersons dataset. By utilizing the CRF optimization, the performance can be improved by 2.3% AP while reducing the miss rate by 2.4%. Thus, the final results show that spatial coherence modeling is effective. The fractal prior improves AP by a further 1.8%, especially helpful for partial gait. Geometric consistency provides a robust feature representation. The combined framework achieves the best performance with 85.2% AP@0.5 and 8.2% LAMR. The only modest reduction in runtime (from 6.8 to 5.6 f/s) can be attributed to the fractal feature extraction and CRF inference, making it acceptable for real time applications.

The optimal hyperparameters found by the grid search and cross-validation are listed in Table 6. The weights \(w_1\), \(w_2\), \(w_3\) specifically balance appearance similarity, spatial smoothness and fractal coherence, respectively. The appearance terms should have higher weight (0.5) since they are more important for differentiating pedestrians, who are mostly similar in appearance and spatial properties. The widths of the Gaussian kernels are denoted by \(\theta_\alpha\), \(\theta_\beta\), and \(\theta_\gamma\) respectively for the spatial and intensity sensitivity of the pairwise interactions. In other words, the different widths of the Gaussian kernels are tuned to the typical scale of pedestrians as found in surveillance imagery. An optimal sensitivity to the geometric variations and robustness against noise is given by the fractal kernel bandwidth \(\sigma_{\text{fractal}} = 1.5\). The weight for the shape complexity (γ) was set to 0.15, and the weights for the pure self-similarity (β) and for the dimension (α) were set to 0.2 and 0.1 respectively. This is done to emphasize the irregularity of the boundary more than internal shapes.

4.3 Ethics

The pedestrian detection is being researched for safety in public surveillance applications. All datasets utilized are publicly available and appropriately licensed for use. The algorithm aims at general pedestrian detection, not demographic classification and identification other than tracking. Local privacy laws and ethical standards must govern deployment of surveillance systems.

5. Discussion

5.1 Mathematical basis for performance improvements

The strong performance of our system is caused by the positive interaction of fractal concepts and mathematical models, creating a dependable method to analyze geometric details and spatial coherence within and between images. Researchers discovered a method for measuring the invariability of human gait as something more than just a scale of time but rather a form of a specific dimension which applies at a different scale. The core of this model relies on the concept of scale invariance within fractals, showing how walking patterns are self-similar. This enables the development of features that prevail despite different levels of technology that are used to develop the video, ultimately solving a critical obstacle in discovering the identity of all people surrounding us daily. The features of self-similarity are used to control human movement in order to not include incoherent and false data. The proposed method's optimization results in a coherent probability, relating to a type of energy principle in statistical mechanics systems. The use of the improved function contains a balance between the features of local and global features as well as the use of global geometric consistency. Utilizing the dual optimization of the energy function allows pixel classification to align with customary fractal properties and to be in spatial coherence by way of pairwise interactions.

The efficacy from the merger comes from the natural harmony between fractal geometry and CRF optimization combined effort itself. Fractal geometry determines the possible positions of the CRF, the image information, to produce realistic and logical scenes. The CRF always runs smoothly between images. Combining facial identification systems with overall body language promotes well-rounded tracking systems. Strong performance in challenging conditions is possible with robust, multi-dimensional analysis.

5.2 Robustness Analysis and Limitation Discussion

Extensive robustness studies led to valuable discoveries about the performance parameters and restrictions of this newly designed method. The test presented an insignificant drop in performance when it comes to occlusion. This happens because the fractals have something to fill in when some of the visual information is lost. Any kind of performance analysis clearly shows that having periods where a person cannot move would break their normal gait cycle, making it inefficient. Studies show that the fractal feature extraction scheme possesses the ability not to be affected by brightness but can have its functionality decreased if the lighting becomes extremely harsh. To calculate the accurate ratio of edges to area in a given image that defines the fractal dimension, low edge contrast causes a limitation as a result, indicating that integrating additional pre-processing methods could improve precision. Fractal dimension calculation slows down the entire inference by 15%. The complexity of an application becomes noticeable in this small fragment of code. These inherent methods, as serial, counteract some of the computational power provided by the GPU for the user. Restrictions are more noticeable when they cause inconvenience on multiple occasions or block a whole operation like in cases of highly complex video analysis. The approach of generalizing makes things better, and a lot of the times, transferability is all about results and outcomes. Although controlled laboratory settings perform fairly well with face recognition algorithms, a deeper analysis suggests that performance quality may increase with the integration of real-world conditions, thereby diminishing the domain discrepancy issue.

5.3 Beyond edges: proposed applications in computer vision and AI

The collaboration of deep learning programs and fractal geometry shows the potential for combining long-established mathematical concepts into modern-day AI technology. With the advent of deep learning has come the feature-learning trend of using data to drive network design. However, no longer will deep learning be dependent on data, as the trend is shifting towards the mathematically-driven and data-free structure and architecture. In this paper, the given proposed idea of Fractal Regularization can possibly act as a model for combining geometric prior knowledge found in our area of study into network design. Furthermore, this method can definitely be adapted to other diverse applications which need to uncover aspects that vary by scale. This expression presents a useful mathematical resource that will guide the encoding of knowledge in any representation. The reasoning on the fractal properties, as they take into account the properties of a geometric space, better informs conventional deep convolutional network systems. Domain specialists can acquire an understanding of the decision-making procedure and express that the results are valid and trustworthy. In certain scenarios like safety and security, the usage of a transparent screen is very favorable.

Real-time performance results clearly show the benefits of utilizing mathematically augmented architectures in practical scenarios. Fractal computing optimizes the techniques which, if applied to edge computing, are advantageous to extending similar geometric enhancements, as edge computing grows in popularity for IoT distributed AI processing.

The proposed methods suggest deeper notions about how the process of computer vision could be further advanced in all areas of application. Researchers have found a new mathematical method to use geometric priors and make predictions on structured formats by combining a probabilistic graphical model with a CRF under fractal regularization. These proposed computational methods incorporate theoretical precision with the practical benefits of speedy operation. These parallel box-counting algorithms and GPU-based realization play a crucial engineering role in the extensive use of fractal-based computer vision systems. The experimental validation approach works with statistics and a confidence range to tell the limitations of a certain performance goal in computer vision. The multi-dataset analysis established a baseline for evaluating the robustness of pedestrian detection so comparative studies can assess its merits. The fusion of age-old mathematical concepts with modern neural network innovation is a methodology that combines classic computer vision techniques with contemporary artificial intelligence practices. While researchers and technologists are increasingly using data-driven solutions to build real-world AI systems, mathematical theory is still necessary too.

6. Conclusion

Today's world has undergone many changes because of many improvements. Surveillance of streets, passages and public places is an important issue that is currently being met by the installation of surveillance cameras. Located in one area, these cameras can record and even analyze any activity of individuals and other objects. Developing analytical systems on these cameras requires smart methods to identify, detect and track targets. This is important because in places such as airports, the behavior of suspicious persons can be investigated, in the streets aggressors can be identified, and other uses as far as possible from reducing crimes and can have beneficial effects on a society. Certainly, the creation of such intelligent systems requires intelligent methods and algorithms that use the principles of artificial intelligence. Therefore, this research provides an intelligent system for detecting pedestrians from videos and then tracking them in different video frames. The proposed approach is initially presented as a pre-processing step, followed by the deep learning techniques with VGG-Net combined with the CRF method, named VGG-Net-CRF, and based on the fractal model for training operations and data testing along with feature extraction. The results represent that real-time pedestrian tracking and detection occurred with higher accuracy and faster execution time than previous methods. The proposed approach tries to recognize patterns for which VGG-Net-CRF is a neural self-organized pattern recognition method. In future research, more magnification and higher quality videos can be used to perform authentication operations with a high-precision proposed approach after detection and tracking.

Frequently asked questions

What is a partial gait cycle, and why is it hard to detect people from one?

A gait cycle is one full stride of walking; a partial gait cycle is only a fragment of that motion — which is what surveillance cameras often capture when a person is briefly in view, occluded, or moving quickly. It is hard because low-quality city cameras typically render pedestrians with only 100 to 200 pixels of target, so classic machine-vision techniques fail and the detector must infer the whole from an incomplete, noisy signal.

How does the VGG-Net-CRF framework work?

A VGG-Net convolutional backbone produces per-pixel pedestrian classification scores that act as unary potentials for a fully connected Conditional Random Field. The CRF enforces spatial coherence through pairwise Gaussian kernels, and a fractal regularization term adds geometric consistency. The network is trained end-to-end with the Adam optimizer and a composite loss that balances classification accuracy with fractal and CRF constraints.

What role does fractal geometry play?

Fractal features exploit the self-similarity of human locomotion across scales, giving a scale- and orientation-invariant measure of silhouette complexity. Used as an auxiliary channel and a regularizer, the fractal prior stabilizes silhouettes under partial gait and occlusion, and in the ablation it improves average precision by about 1.8% in the partial-gait regime.

Does the method run in real time?

Yes. The full framework processes video at 5.6 frames per second — about 0.18 seconds per frame — on a single RTX-class GPU, using mean-field CRF inference with a fixed five-iteration count, GPU-accelerated fractal computation, feature caching, structured pruning that removes 15% of connections, and 8-bit quantization of feature maps.

References

1. Zafra-Palma, J., Marín-Jiménez, N., Castro-Piñero, J. et al. Health & Gait: a dataset for gait-based analysis. Scientific Data 12, 44 (2025).
2. Viola, Paul, Jones, Michael, and Snow, Daniel. 2003. Detecting Pedestrians Using Patterns of Motion and Appearance. Mitsubishi Electric Research Laboratories.
3. Hariyono, Joko, and Jo, Kang-Hyun. 2016. Detection of Pedestrian Crossing Road: A Study on Pedestrian Pose Recognition. Neurocomputing, Accepted Manuscript.
4. Wang, Yi, Piérard, Sébastien, Su, Song-Zhi, and Jodoin, Pierre-Marc. 2016. Improving pedestrian detection using motion-guided filtering. Pattern Recognition Letters, pp. 1–7.
5. Ribeiro, David, Nascimento, Jacinto C., Bernardino, Alexandre, and Carneiro, Gustavo. 2016. Improving the performance of pedestrian detectors using convolutional learning. Pattern Recognition, Vol. 61, pp. 641–649.
6. Liu, Yazhou, Lasang, Pongsak, Siegel, Mel, and Sun, Quansen. 2016. Multi-sparse descriptor: A scale invariant feature for pedestrian detection. Neurocomputing, Article in Press.
7. Chen, Yanxiang, Zhang, Luming, Liu, Xiao, and Chen, Chun. 2016. Pedestrian detection by learning a mixture mask model and its implementation. Information Sciences, Vol. 372, pp. 148–161.
8. Qi, Bin, John, Vijay, Liu, Zheng, and Mita, Seiichi. 2016. Pedestrian detection from thermal images: a sparse representation based approach.
9. Mateus, André, Ribeiro, David, Miraldo, Pedro, and Nascimento, Jacinto C. 2019. Efficient and robust Pedestrian Detection using Deep Learning for Human-Aware Navigation. Robotics and Autonomous Systems, Vol. 113, pp. 23–37.
10. Li, Chengyang, Song, Dan, Tong, Ruofeng, and Tang, Min. 2019. Illumination-aware faster R-CNN for robust multispectral pedestrian detection. Pattern Recognition, Vol. 85, pp. 161–171.
11. Zhang, Shan, Yang, Xiaoshan, Liu, Yanxia, and Changsheng, Xu. 2020. Asymmetric Multi-stage CNNs for Small-scale Pedestrian Detection. Neurocomputing, in press, available online 26 May 2020.
12. Ben Khalifa, Anouar, Alouani, Ihsen, Mahjoub, Mohamed Ali, and Essoukri Ben Amara, Najoua. 2020. Pedestrian detection using a moving camera: A novel framework for foreground detection. Cognitive Systems Research, Vol. 60, pp. 77–96.
13. Lu, Ruiqi, Ma, Huimin, and Wang, Yu. 2020. Semantic head enhanced pedestrian detection in a crowd. Neurocomputing, Vol. 400, pp. 343–351.
14. Ding, Lu, Wang, Yong, Laganière, Robert, Huang, Dan, and Fu, Shan. 2020. Convolutional neural networks for multispectral pedestrian detection. Signal Processing: Image Communication, Vol. 82, March 2020, Article 115764.
15. Wei, Xing, Zhang, Haitao, Liu, Shaofan, and Lu, Yang. 2020. Pedestrian detection in underground mines via parallel feature transfer network. Pattern Recognition, Vol. 103, July 2020, Article 107195.
16. Zou, Tengtao, Yang, Shangming, Zhang, Yun, and Ye, Mao. 2020. Attention guided neural network models for occluded pedestrian detection.