Study on Slam Algorithm Based on Object Detection in Dynamic Scene

In order to reduce the deviation of track estimation and scene reconstruction errors caused by unreliable feature points extracted from dynamic objects in dynamic scenes, a visual SLAM algorithm based on object detection is presented in this paper. It detectes dynamic objects via Yolov3 which is a deep learning algorithm, and a point cloud map for scene reconstruction is generated by ORB-SLAM2 finally. To reduce the effect of detection missing, sliding window compensation algorithm is presented. Empirical evaluation on the FR3 series of TUM dataset demonstrates the effectivenessof our approach: 1) most scenes are reconstrusted in point cloud maps generated by our algorithm, and 2) The average decrement of exception frames reaches to72.97% and key frames decrease 57.34% when compared to ORB-SLAM2without our compensation algorithm, and 3) the average running time of the whole process is about 150ms, which basically meets the real-time requirements.