Abstract: Aiming to solve the problem of local obstacle avoidance for unmanned surface vehicle under complex marine conditions，deep learning was introduced to deal with the vision information and completed the local path planning for unmanned surface vehicle combining with VFH + algorithm． The image semantic segmentation model with symmetrical encoder-decoder structure and faster RCNN model were used to detect water edge and identify obstacles on the surface of the water to build environment model around unmanned surface vehicle． Then local path planning method based on VFH + was brought to construct the primary histogram，binary histogram and mask histogram to compress the environment data，and a reasonable cost function was introduced to obtain the effective obstacle avoidance direction for unmanned surface water． Results of simulation experiments based on MODD dataset and obstacle avoidance experiments in real ship show that the algorithm can effectively extract the water surface image information and obtain a reasonable local path planning strategy to lead a smooth obstacle avoidance path at 10kn， which meets the obstacle avoidance requirements of unmanned surface vehicle． 

Key words: unmanned surface vehicle, image segmentation, faster RCNN model, VFH + algorithm, path planning