PTZSort: Using Reinforcement Learning to Drive PTZ Camera for Vehicle Detection
Runze Zhang · 2024
In the evolving field of traffic management and intelligent transportation systems, the technology for vehicle identification and tracking plays a pivotal role. Traditional methods often fall short in complex scenarios due to scene intricacy and target occlusion, leading to decreased recognition and tracking accuracy. This study proposes a new method called PTZSort, which utilizes Pan Tilt Zoom (PTZ) cameras that come with advanced pan-tilt and zoom functionalities. To enhance the functionality of PTZ cameras further, we propose a reinforcement learning-driven algorithm for automating tracking processes. This innovation allows the camera to adaptively learn from its environment, optimizing tracking behavior without manual intervention. We also integrate this with an advanced version of the DeepSort algorithm, refining vehicle tracking efficiency by effectively managing occlusions and dynamic scene changes. Moreover, the inclusion of a Squeeze-and-Excitation Network within PTZSort significantly bolsters the system's robustness against interference and enhances its ability to extract salient features from target vehicles. The design of this network architecture emphasizes adjusting channel-specific feature responses by directly modeling the relationships between channels. This approach enhances the distinctiveness of traffic management and security systems.