Real-time UAV-based vehicle counting method based on deep learning

• Main Author: Dang, Zichen
• Other Authors: Yap Kim Hui
• Format: Thesis-Master by Coursework
• Language: English
• Published: Nanyang Technological University 2024
• Subjects: Computer and Information Science; Object counting; FasterNet; Attention module; YOLOv7; DeepSORT; UAV
• Online Access: https://hdl.handle.net/10356/180873
• Institution: Nanyang Technological University

Vision-based vehicle counting is essential in the traffic field. By utilizing intersection surveillance cameras, Unmanned Aerial Vehicles (UAVs), or existing filming devices, vision-based vehicle detection and counting can provide reliable traffic flow information to traffic control systems. Among these methods, UAV-based vehicle counting has gained significant attention due to its high flexibility, wide field of view, and fast deployment capabilities. Typically, real-time video vehicle counting methods are employed to reflect traffic flow data instantaneously. However, deep learning-based vehicle detection and counting still face challenges related to efficiency and accuracy. To address these issues, this dissertation proposes a video vehicle counting method that integrates an improved YOLOv7 with the DeepSORT tracking algorithm. In the design of the detector, the high altitude of images captured by UAVs typically results in smaller vehicle sizes, while the traffic scene often presents a complex background. Additionally, real-time video detection demands processing speeds. To address these challenges, this dissertation proposes an improved YOLOv7 detector. The detector uses FasterNet as its backbone network and replaces traditional convolution with partial convolution, effectively reducing computational redundancy and minimizing memory access, thereby improving detection speed. Furthermore, a Convolutional Block Attention Module (CBAM) is integrated into the feature fusion network to improve feature discrimination ability in complex environments, thus improving both detection accuracy and robustness across various traffic scenarios. The experimental results demonstrate that the improved detector outperforms traditional networks in terms of detection accuracy and processing speed. In the counting algorithm, a tracking-by-detection strategy is employed, combining the improved YOLOv7 detector with the DeepSORT tracker to effectively track and count vehicles. The method determines the counting area by delineating a judgment line, and determines the number of vehicles passing through the intersection by checking if the tracking trajectory crosses this line. The experimental results indicate that the proposed counting model demonstrates significant advantages in both real-time processing speed and counting accuracy when compared to the baseline model.