A Real-time Algorithm for Visual Detection of High-speed Unmanned Surface Vehicle Based on Deep Learning

We proposed a high-robustness real-time visual detection algorithm based on deep learning, which is aiming at the problem that high-speed unmanned surface vehicle(USV) are difficult to detect and identify targets in complex mission scenarios. We design feature extraction network based on MobileNet, by removing average pooling layer, full connectivity layer and softmax layer. We then add 8 convolution layers to improve feature extraction capability. Additionally, we built SSD structure detection network to achieve fast multi-scale detection in fuse selected Multi-size feature map results. In the end, we implement and test algorithm on the embedded GPU. The results show that our algorithm can detect multiple types of specific targets on water in real time, with strong robustness and multi-scale characteristics. The detection time of single-frame video can be completed within 50ms. Through our video simulation experiments, the algorithm has high detection rate and strong robustness to the actual detect situation, while has important engineering practical value.