FCN and Siamese Network for Small Target Tracking in Forward-looking Sonar Images

In underwater forward-looking sonar images, a small moving target is susceptible to noise pollution, and the performance of object tracking is greatly affected by background disturbances, illumination changes and occlusion. Hence, we propose to combine FCN and Siamese network for small moving target tracking. In order to solve the problem of too few data sets, we use geometric transformation methods to extend the data sets. In the other side, we adopt the FCN network structure, it can accept any size of input forward-looking sonar images and make tracking more efficient. Moreover, by using the Siamese network structure and removing the last full connected layer, it enables tracking more accurately. The reduction in the number of network layers also greatly improves real-time performance. The experimental results show that our method is very suitable for small moving target tracking in forward-looking sonar images and there is no target tracking loss occurred. It overcomes the noise interference in forward-looking sonar images, and significantly improves the accuracy and real-time performance.