Convolutional neural network based obstacle detection for unmanned surface vehicle

Unmanned surface vehicles (USV) is the development trend of future ships, and it will be widely used in various kinds of marine tasks. Obstacle avoidance is one key technology for autonomous navigation of USV. Convolutional neural network based obstacle classification and detection method is applied to USV visual images in environment sensing task. To solve the problem of low detection and classification accuracy of obstacles in the visual inspection of USV, a bidirectional feature pyramid networks is proposed combining hybrid network architecture of ResNet and improved DenseNet. The proposed method can further enhance the detection and classification some types of obstacles by using the underlying multi-layer detail features and high-level strong semantic features in the network architecture. The detection and classification performance of the proposed method is evaluated on a self built dataset. Ablation experiments and performance tests on open datasets are also employed. The experimental results show that the proposed algorithm has best performance for obstacles detection, and it is more suitable for autonomous navigation of USV.