An Adaptive Feature Extraction Visual SLAM Method for Autonomous Driving

Due to the traditional visual Simultaneous Localization and Mapping(SLAM) method is easily affected by illumination, and the uneven distribution of feature points leads to a decrease in localizing accuracy or even tracking failure, we propose a visual SLAM method for adaptive ORB feature extraction and division based on ORBSLAM2. First,we dynamically set the number of extracted feature points for each frame to enhance the stability and robustness of tracking and reduce the occurrence of tracking failures. After that, we design a FAST feature point extraction strategy with a local dynamic threshold. Compared with the fixed threshold method, this can reduce the impact of illumination changes on feature point extraction. Finally, we adaptively limit the number of quadtree node division layers according to the expected number of feature points to avoid local over-concentration. We test our method on the KITTI dataset and the automatic guided vehicle platform. Compared with ORBSLAM2, the feature points extracted by our method are more uniform, and the localizing accuracy and robustness are improved.