Siamese Network Using Adaptive Background Superposition Initialization for Real-Time Object Tracking

Object tracking has become widespread in many fields, such as autonomous vehicles, video surveillance and robotics. However, it is far from the requirements for real-world applications. Recently, Siamese network based trackers have attracted high attention by balancing accuracy and speed. Because these trackers only learn a similarity measurement model via off-line training, the exemplar branch has insufficient discriminant information to adapt to the constantly changing appearance of the target in subsequent frames. We propose a Siamese network based tracker that improves upon tracking performance as follows. First, an adaptive background superposition initialization is proposed and used in the exemplar branch to make full use of the limited prior information in the first frame. Second, a light-weight convolutional neural network is proposed and applied as the tracker's backbone; it compresses the dimensions of the feature to ensure speed and accuracy. Third, the channel attention module is introduced into our tracker and integrated with adaptive background superposition initialization. The feature map of the original exemplar image and its background changed image are adjusted by a channel attention model and fused to enhance the representation of the exemplar image. The GOT-10k dataset is applied to train our tracker. Finally, experiments on the object tracking benchmark (OTB) and visual object tracking (VOT) demonstrate the effectiveness of our proposed approach compared with state-of-the-art trackers.