Dynamic feature point removal method with joint target detection and depth information

Aiming at the problem that the localization accuracy and robustness of visual localization systems are easily affected by dynamic feature points in dynamic environments, a dynamic feature point removal method combining target detection and depth information is proposed. The YOLOv7 target detection network is introduced to quickly obtain the target category and position information of the current image frame, and the coordinate attention (CA) mechanism is added to optimize the deep learning model and improve the target detection accuracy of the network. In addition, a dynamic feature point optimization strategy using depth information and pairwise geometric constraints is proposed. Dynamic feature points are effectively eliminated while as many static points as possible are retained, thus reducing the impact of dynamic points on the localization accuracy and robustness of the system. Experimental validation is performed on the publicly available dataset TUM. The results show that the proposed scheme has obvious advantages in terms of localization accuracy and robustness compared with ORB-SLAM2. At the same time, compared with DynaSLAM, the localization accuracy is basically the same, but the operation speed is significantly improved.