PIDIVO:A Visual Odometry Combining PiDiNet and Adaptive Binarization Strategy

In low-texture environments, visual odometry often suffers from reduced tracking accuracy and loss due to insufficient point feature extraction. To address this issue, a new method is proposed that applies PiDiNet to visual odometry, called PiDiVO. By combining PiDiNet with an adaptive binarization strategy, robust edges can be extracted from images, effectively overcoming the problem of insufficient point feature extraction in low-texture environments. Meanwhile, a reasonable keyframe selection mechanism reduces cumulative drift in visual odometry, thus improving its robustness. In addition, determining keyframes based on tracking quality ensures stable tracking during drastic image changes, and periodically selecting keyframes helps reduce visual odometry drift. We evaluated PiDiVO on the TUM RGB-D and ICL-NUIM datasets and compared it with state-of-the-art methods such as ORB-SLAM2, CannyVO, and REVO. The results show that PiDiVO demonstrates better robustness in low-texture environments and under illumination changes, while also being capable of real-time operation.