Obstacle Detection of Unmanned Surface Vessel based on Faster RCNN

With the development of unmanned surface vessel in the world today, obstacle avoidance using environmental information is the basis to ensure its high maneuvering performance and safety. However, directly using standard algorithms will lead to missing and wrong identification severely for characteristics of marine obstacles. This paper adds a multi-scale feature extraction layer of dilation convolution and group convolution to Faster Region based Convolutional Neural Network (Faster-RCNN), the baseline model, and changes the classification algorithm to improve its robustness and accuracy. Soft-Non Maximum Suppression (Soft-NMS) is used to enhance the prediction effects further. After improvements, the mean average precision value increases by 3.35%, and the final loss value decreases by 0.20. Given the phenomenon of missing and misidentification in the prediction by the baseline model, the results of our new model show outstanding performance.