Visual odometry (VO) systems estimate the real-time pose of robots, a fundamental challenge in the field of mobile robotics. Among various sensor configurations, cameras are widely adopted due to their low cost, low power consumption, and ease of integration. However, visual systems are inherently fragile, with performance significantly affected by illumination factors such as lighting variations and exposure conditions, resulting in instability in complex environments.

This thesis addresses the robustness challenges of VO systems under high motion and high dynamic range conditions by focusing on camera exposure control algorithms. Leveraging the principles of camera imaging, we calibrate the relationship between scene illumination and pixel intensities. Utilizing the bracketing functi...[ Read more ]

Visual odometry (VO) systems estimate the real-time pose of robots, a fundamental challenge in the field of mobile robotics. Among various sensor configurations, cameras are widely adopted due to their low cost, low power consumption, and ease of integration. However, visual systems are inherently fragile, with performance significantly affected by illumination factors such as lighting variations and exposure conditions, resulting in instability in complex environments.

This thesis addresses the robustness challenges of VO systems under high motion and high dynamic range conditions by focusing on camera exposure control algorithms. Leveraging the principles of camera imaging, we calibrate the relationship between scene illumination and pixel intensities. Utilizing the bracketing functionality of modern camera sensors, we propose three novel exposure control algorithms tailored to distinct system frameworks: In the initial segment, we enhance the traditional feedback-based exposure control by introducing image synthesis techniques, decoupling parameter estimation from hardware feedback. This approach significantly improves parameter update rates from 20 Hz to 200 Hz, enhancing system responsiveness to quick illumination change. In the subsequent segment, we propose a one-shot global optimal exposure parameter algorithm, employing bracketing techniques to capture multi-exposure representations of a scene. By regressing the relationship between image quality metrics and exposure parameters, this method bypasses traditional feedback control and directly determines optimal exposure settings. In the third segment, we design an intelligent exposure control agent based on deep reinforcement learning. Through an offline training framework and customized reward functions, the agent learns to adapt exposure parameters proactively, mitigating catastrophic failures by considering motion dynamics and lighting distributions. The proposed methods fully exploit the robustness and adaptability of visual sensors in complex scenarios, providing stronger reliability and performance support for visual odometry tasks.