Real-time tracking of visual objects based on deep feature fusion

In order to improve SiamFC's recognition and localization capabilities in complex scenes, this paper proposes a visual object tracking network structure based on deep feature fusion. Aiming at the recognition ability, the feature extraction network IVBRNet with deep structure is designed based on VGGNet and ResNet, and the deep features with rich semantic information extracted by IVBRNet are used to assist object recognition. For the positioning ability, the ECA attention module is added at the end of the original AlexNet object feature extraction branch and the original Relu activation function is replaced with the ELU activation function to improve the expression ability of the object feature. Finally, the response graphs obtained by AlexNet and IVBRNet are calculated and fused to perform object tracking together. The results on the OTB2015 and VOT2016 data sets show that the algorithm has a more stable tracking effect than other algorithms, and can better cope with scenes such as fast moving and complex backgrounds. At the same time, the algorithm reaches 32 frames per second, which meets real-time requirements.