Research On Deep Learning For Object Tracking

Object tracking is a classic and important research direction of computer vision,and it is widely used in various fields.Its main task is to continuously track the position of the target in subsequent frames given the target object in the first frame.Object tracking is a challenging research topic,because it needs to deal with difficult factors such as occlusion,deformation,and illumination variation,so it has been widely concerned by scholars for many years.In recent years,with the development of deep learning technology,the fully convolutional Siamese networks have shown great potential in the field of object tracking.In particular,the tracking algorithm SiamFC can well balance speed and accuracy.At the same time of high tracking accuracy,it can also meet the real-time requirement.However,the tracking results of the algorithm in some scenarios are not ideal and it is prone to drift,especially when similar distractors appear in the background.This paper improves the algorithm according to its problems and enhances the robustness and accuracy of the algorithm.Based on SiamFC,this paper proposes an algorithm SiamMFF based on fully convolutional Siamese networks and multi-level feature fusion structure.Different layers in convolutional neural networks contain features of different properties.Low-level layers include fine-grained spatial information,and high-level layers incorporate abstract semantic information.Therefore,it is useful to fuse features from different levels of the network to improve the discriminative ability.SiamMFF uses a multi-feature fusion structure to fuse the features from the first,third,and fifth layers of the network.Hence,the feature representations it extracts can not only distinguish objects from different categories,but also differentiate target from similar objects,and it is able to track stably in complex scenes.This paper compares SiamMFF with nine other excellent object tracking algorithms on the OTB benchmark.SiamMFF achieves high accuracy and success rate as well as better performance.In order to improve the robustness and accuracy of the algorithm,this paper proposes an object tracking algorithm TripMFF based on Triple network and multi-level feature fusion structure.The algorithm adopts Triplet network as backbone architecture,and takes the first frame,the previous frame,and the current frame of the image sequence as the network input,which makes full use of the entire video and can better adapt to the target variations.Furthermore,the network also incorporates features from multiple layers,thereby extracting more discriminative feature representations.In addition,TripMFF utilizes the APCE-based response map fusion strategy to locate the target more accurately by judging the confidence of the response maps.The paper compares TripMFF with some other excellent trackers on the OTB and VOT benchmarks.TripMFF significantly improves the tracking performance on the basis of the original algorithm,and obtains leading tracking results,proving its advancement.