Accurate Positioning Siamese Network for Real-Time Object Tracking

Though end-to-end siamese networks have achieved great performance on object tracking owing to offline pre-training with large datasets. They are still liable to fail to track fast moving object and their accuracy suffers from the cosine window for mitigating background interference. The cosine window will aggravate the boundary effect and have a negative impact on track. In this paper, we propose an accurate positioning siamese network (FPSiam) for real-time object tracking. This approach can better estimate the search candidate area of the object in the current frame from the previous tracking position and can keep tracking well for the fast moving object. The unbalanced penalty mechanism of the cosine window in the tracking process is avoided. Furthermore, the features of the low-level layers are high-resolution, which is suitable for positioning the object. High-level layers are full of rich semantic features which are suitable for classifying the objects. In order to utilize the advantages of high-level features and low-level features, we introduce a densely connected tracking network. A number of experiments were conducted on five challenging tracking datasets: OTB50, OTB2013, OTB2015, VOT2015, and VOT2016, and the proposed method achieved excellent results on these benchmarks.