Real-time object tracking in the wild with Siamese network
Feng Han,Shaokui Jiang,Jianmin Wu,Baile Xu,Jian Zhao,Furao Shen
DOI: https://doi.org/10.1007/s11042-023-14519-6
IF: 2.577
2023-02-18
Multimedia Tools and Applications
Abstract:Single object tracking (SOT) is one of the most important tasks in computer vision. With the development of deep neural networks and the release for a series of large scale datasets for single object tracking, Siamese networks have been proposed and perform better than most of the traditional methods. However, recent Siamese networks are getting slower to obtain better performance as they become deeper. Most of those networks could only meet the needs of real-time object tracking in ideal environments. In order to achieve a better balance between efficiency and accuracy, we propose a simpler Siamese network for single object tracking, which runs fast in poor hardware configurations while remaining an acceptable accuracy. The proposed method consists of three parts: sample generation, SE-Siamese and regression localization. In the sample generation stage, template patch and detection patch are cropped from the selected video frames in a new way. The SE-Siamese subnetwork adopts Siamese network and Squeeze-and-Excitation (SE) network as the feature extractor which is an effective way of speeding up the training phase. The regression localization network aims to compute the location of the tracked object in a more efficient way without losing much precision. To validate the effectiveness of the proposed approach, we conduct extensive experiments on several challenging tracking benchmark datasets, including VOT2015, VOT2016, VOT2017 and OTB-100. The experimental results show that our approach displays significant speed improvements compared to several strong baseline trackers (19.5 FPS to 44.4 FPS).
computer science, information systems, theory & methods,engineering, electrical & electronic, software engineering