Object Detection under Dynamic Environment with Application in Visual SLAM

In recent years, visual SLAM has been applied to more complex environments. However, the traditional visual SLAM algorithms cannot meet the practical requirements in dynamic environments. In this paper, we propose a improved dynamic object detection algorithm via a trained NCNN-based MoblienetV3 network, where the dynamic feature points in the bounding box are determined by using epipolar constraints, such that the dynamic feature points in an image are tracked and built through the KLT optical flow methods. Finally, the proposed dynamic object detection and tracking algorithm is applied in visual SLAM, where the trajectory estimation errors are minimized via the local mapping thread and the accumulated errors are eliminated in the loop closing module. By comparing with ORB-SLAM2 via the TUM dataset, the effectiveness of the proposed algorithm has been verified.