Crowd-based people detection using deep learning

• Main Author: Chen, Lei
• Other Authors: Yap Kim Hui
• Format: Thesis-Master by Coursework
• Language: English
• Published: Nanyang Technological University 2022
• Subjects: Engineering::Electrical and electronic engineering
• Online Access: https://hdl.handle.net/10356/155650
• Institution: Nanyang Technological University

Human detection is a hot research field in computer vision. Its purpose is to detect humans in images or video sequences and provide precise location by using bounding boxes. One major challenge is to detect people in crowded environments. Specifically, this challenge is broken into two main difficulties: multi-scale problems and occlusion problems. Most object detection algorithms have low accuracy for small object detection and identification. When objects are occluded, the features will overlap, making it hard to determine a correct outcome and leading to false or missed detection. This project first reviewed an extensive list of literature related to object detection based on handcraft and deep learning methods. Then, two state of art neural networks were introduced (EfficientDet and YOLOv5), and through further analysis, I analyzed the components and the thesis and the subsequent source codes, deduced the complete network structure, and explained the specific implementation process of the critical parts. From this, I used the Crowd- Human data set to train the aforementioned networks, and got two object detection networks for people in crowded environments. Experimental results suggest that compared to Faster RCNN, EfficientDet yields detection speeds three times faster in crowded environments, whereas YOLOv5 provides the fastest detection speed while increasing AP by almost 18%. In addition, YOLOv5 solves the difficulty of detecting small-scale objects while maintaining high accuracy of objects in crowded environments. Overall, YOLOv5 can effectively detect people in crowded environments and prevent missed detection of hard-to-detect small-scale people. Its high detection accuracy and fast detection speed make an effective option for crowded environment scenarios. Keywords: Deep Learning, Object Detection, Human Detection.