Anchorless target tracking based on multilayer feature fusion

Target tracking is an important part of the field of computer vision, widely used in systems such as unmanned aerial vehicles, automatic driving, sports events, etc., and is closely related to target detection, image segmentation and so on. Aiming at the traditional twin network-based target tracking algorithm which is often prone to tracking failure when facing complex scenes such as object's deformation, scale change, occlusion, etc., this paper proposes a multilayer feature fusion method for fusing deep features and shallow feature information, uniting the feature maps of the multilayer network, composing a continuous feature space, and fusing the high-latitude features of the deeper network under the premise of preserving the details of image features, which enriches the image feature information extracted by the network model, improves the diversity of feature extraction, and also reduces the number of parameters for network training. Experimental results on the VOT dataset show that this method is higher than SiamFC, SiamVGG, DSiam and other methods in tracking accuracy.