A Real-time Long-term Object Tracking Algorithm Based On Deep Learning

The real-time long-term target tracking algorithm based on deep learning designed in this paper is mainly to solve the problem that the image-based target tracking algorithm is not good in real-time and poor in robustness when used in industry.The paper uses the convolutional features extracted by deep learning to replace the traditional features,which ensures the effectiveness of the algorithm.Using the triplet loss instead of the contrast loss and modifying the template update method ensure the long-term performance of the algorithm.Using model simplification and algorithm acceleration ensure the real-time performance of the algorithm.Finally,we implement a long-term,real-time and robust object tracking algorithm for real-world scenarios.The main contributions and innovations of this paper are as follows:1.In the convolution feature extraction training,the triplet loss(Triplet Loss)is introduced.Compared with the features obtained by using the contrast loss training,more attention is paid to distinguishing the differences between different categories.,improve the quality of the features.Compared to the direct use of pre-trained convolution features without Fine-Tuning,features are more concerned with the deformation of objects.The dimensions of the feature are increased compared to traditional image features.These three improvements have improved the final accuracy and accuracy,and also solved the problem of manually adjusting the threshold due to the common features of convolution extraction.2.In the data set preparation process,random cropping is introduced to increase the number of positive samples.Since the target(positive sample)in the video usually has only one frame per frame,but the background is very large,the target frame of each frame is enlarged and then randomly cropped as a positive sample,which not only increases the number of positive samples,but also equalizes the samples.It also increases the generalization ability of the algorithm.3.When the target tracking algorithm is executed,always select the first frame as the template and search for the target within the entire video frame.Compared with the majority algorithm,the previous frame is used as the target,and only the target is searched near the previous frame target,which ensures the long-term performance of the algorithm,can track the target that disappears after disappearing,and solves the problem of high-speed moving target.4.Increased data-driven model simplification,eliminating unnecessary nodes and branches,reducing the number of operations.When the target tracking algorithm is executed,we need to calculate the convolution feature once every frame of the video.At the same time,in order to ensure the long-term performance of the algorithm,the entire video frame is required to participate in each iteration,which will result in slow operation.Therefore,the model is simplified to achieve model acceleration,which ensures the real-time performance of the algorithm.Finally,the test and verification show that the target tracking algorithm TripNet achieves better accuracy and accuracy than most related filtering algorithms,and the deep learning algorithm is in the middle and upper reaches,that is,the effect of the algorithm can be guaranteed.In the execution speed of the algorithm,it is at the upstream level of the target tracking algorithm,and the FPS is 27 frames per second,which can reach real time.In the long-term test of the algorithm,the success rate is higher than most related filtering algorithms and deep learning algorithms,and the long-term tracking ability is the strongest.