Pedestrian recognition based on deep learning

• Main Author: Gu, Qiwei
• Other Authors: Teh Kah Chan
• Format: Thesis-Master by Coursework
• Language: English
• Published: Nanyang Technological University 2024
• Subjects: Computer and Information Science; Engineering
• Online Access: https://hdl.handle.net/10356/181885
• Institution: Nanyang Technological University

Due to the rapid development in deep learning, vehicle and pedestrian recogni- tion has become increasingly widespread and is now a significant part of computer vision field. Traditional vehicle recognition methods are slow and less accurate. While YOLO-based algorithms have improved accuracy, they still face the challenge of large parameter sizes and slower speeds. From R-CNN to SSD to YOLO algorithms, each of them has its strengths and weaknesses. YOLOv4 stands out for its balance between speed and precision. Considering the diverse requirements of different scenarios for recognition accuracy and speed, this dissertation proposes three improved YOLOv4- based algorithms tailored for different use cases of vehicle and pedestrian recognition. Firstly, the mAP achieved by YOLOv4 on the PASCAL VOC 2007 dataset still leaves room for improvement, making it insufficient for high-accuracy demands such as autonomous driving. To address this, the loss function and feature fusion network were modified. By substituting the CIoU loss function with SIoU, the accuracy of bounding box regression was enhanced. Improving the fusion of shallow features boosted the model’s performance in detecting small objects. Experimental results in- dicate that the enhanced algorithm led to a 2.2% increase in mAP on the BDD dataset. Secondly, although YOLOv4 significantly improves accuracy compared to tra- ditional methods, it has a large parameter size and deep network layers, resulting in slower speeds and higher hardware requirements. To improve this, a modified YOLOv4 algorithm was proposed by replacing standard convolutions with depthwise separable convolutions, reducing redundant parameters and network depth. These were integrated into the ResBlock structure to preserve gradient information. Experiments showed that this method increased recognition speed by about 70%, with an FPS im- provement of 30.9 on the test platform. Although mAP decreased by 3.2%, the model maintained good performance on small object detection. Lastly, for scenarios with lower accuracy requirements and simpler targets, a lightweight algorithm was developed. By applying depthwise separable convolutions and introducing a channel reduction factor, the network became thinner. The YOLO Head was simplified to use only two detection layers, and the feature fusion network was streamlined. Tests with real-world videos showed the model had good recognition accuracy for most vehicles and pedestrians. The key advantage of this algorithm is its low hardware resource requirements, achieving real-time detection speeds and meeting the needs of scenarios where accuracy and computational power demands are low.