Camera Attributes Control for Visual Odometry With Motion Blur Awareness

The practical application of visual odometry (VO) is challenging under High Dynamic Range (HDR) environments which has a significant impact on the quality of feature point tracking. There are two key attributes that affect the image quality of a camera, namely exposure time and gain. The goal of camera attributes control research is to design algorithms that can actively adjust camera acquisition attributes to obtain clear, information-rich images. We propose an efficient and stable camera attributes control method with distinct image quality metrics and a linear convergence search algorithm. Specifically, we propose to use the weighted sum of image gradient and entropy to represent the image quality. Its first order derivative is also computed for a line search algorithm to determine the optimal camera exposure. Then, the algorithm estimates scene change speed by optical flow and finds out the maximum exposure time without motion blur. Ultimately, the maximum exposure time and optimal camera exposure are used to determine the camera properties including exposure time and gain. Additionally, the proposed method speeds up the convergence of the algorithm with the help of simulated image generation which enables nearly optimal attributes to be obtained directly from an overexposed or underexposed image. To validate the proposed approach, we test it in challenging HDR environments with fast motion and compare it with built-in automatic exposure. It will be open source for the benefit of the community.