An Improved Tracking Algorithm Based on MDNet

Abstract

Abstract The MDNet algorithm works well on tracking problems of the video sequence, but the speed is very slow. We have made some improvements to accelerate the process of feature extraction. It enhances the expressive ability of the feature map by removing the max pooling layer and using the method of expanding convolution to increase the receptive field of each point on the feature map. In addition, MDNet is building on CNN, and there are problems that similar targets have a large interference to the results. To address this problem, We use RNN to capture the long-term dependence of the target before and after the target data in the sequence data, and introduce the RNN to model the structure information of the target object, and then fuse the RNN feature and CNN feature of the tracked target object. In addition, another new loss term is introduced to make the targets in different domains away from each other in the shared feature space, thereby improving the algorithm’s ability to discriminate similar interferers. Compared with MDNet, our improved algorithm is much faster and the accuracy is improved.