COMPARATIVE ANALYSIS OF PEDESTRIAN DETECTION USING DEEP CNN, R-CNN, FAST R-CNN AND FASTER R-CNN

Abstract

In the field of computer vision, pedestrian detection is an important job that is fre quently utilised in robots, auto-navigation, and video surveillance. Pedestrian detection is critical to lowering accident rates and enhancing automation in the transportation sector. Deep learning has revolutionised object identification in recent years, inspiring the devel opment of many architectures for this use. In-depth comparisons of four well-known deep learning models for pedestrian detection are presented in this thesis: As an illustration, there are four types of convolutional neural networks: region-based (R-CNN), fast region based (Fast R-CNN), faster region-based (Faster R-CNN), and deep convolutional (Deep CNN). This work’s primary goal is to evaluate these models’ performance on a number of performance metrics, including sensitivity to various situations, temporal complexity, and detection rate. Because each model has a unique way of enhancing efficiency, they are all important milestones in the evolution of object detection architecture. Multiple tests on a standardized pedestrian detection dataset will be included in this thesis to offer a clear grasp of the variations and similarities between these models. It also considers a variety of dynamically changing factors, including pedestrian density, occlusion occurrence, and illumination conditions. The precision, recall, F1-score, and average precision are among the metrics used to evaluate the detection task’s accuracy. It also covers how long each model takes to train and infer, as well as how many resources it uses overall. The re sults show how well the four classes of models performed in terms of accuracy and time. Consequently, Deep CNNs and R-CNNs highlight significant facets of feature extraction and region-based detection techniques, even if Faster R-CNN keeps a favorable accuracy and speed ratio throughout the studies. Though it is somewhat less efficient than Fast R-CNN, which was designed to reach near real-time performance, Faster R-CNN performs better than both R-CNN and Fast R-CNN. This thesis also provides the conclusion for the overall findings and discusses the future scope of this research and future possibilities for the development of pedestrian detection iv systems. The following goals for future research can be identified: introducing the teach ing and using the features of several architectures to improve the effect; strengthening the detection under different environments; expanding the study for other relevant jobs in computer vision. From this comparative study, the authors have provided a rich set of lessons that can be used in the continuous improvement of pedestrian detection and can aid the improvement of numerous intelligent systems in different application areas.