Adaptive Zoom Active Object Tracking Based on Reinforcement Learning

Active Object Tracking (AOT) involves the tracker using visual observations from the camera to keep the object centered in the frame by controlling the motor system. In contrast to extensively researched passive tracking, which primarily focuses on perception, AOT integrates both perception and control tasks, making it more suitable for real-world tracking applications. AOT has the capability to interact with the environment, where perception determines the control actions, and the outcomes of these actions, in turn, shape future inputs and influence perception. We employed a pan-tilt-zoom (PTZ) camera as the active tracker for highly maneuverable targets. By intelligently adjusting the zoom level based on the target's movement, our tracker optimizes the size of the target in the visual input, thereby enhancing the tracking success rate. Recognizing the complex interaction between perception and control, we applied deep reinforcement learning (DRL) to jointly train the tracker's perception and control modules in a virtual environment. The experimental results in the virtual environment demonstrate that our proposed method achieves a higher tracking success rate and speed compared to traditional algorithms.