Research Of Object Tracking Algorithm Based On Deep Learning

As one of the important research topics in computer vision,visual object tracking has a wide application prospect in automatic driving,intelligent traffic monitoring and human-computer interaction.However,the uncertainty of the tracked object and the complexity of the tracking scene bring great challenges to the object tracking task.In practical applications,the real-time and high efficiency of the tracking algorithm should also be considered.In order to obtain a tracking algorithm with better real-time and robustness under a variety of interference factors,this thesis does research on object tracking algorithm based on the difficulty of object tracking.Aiming at the robustness of object tracking algorithm based on full convolution Siamese network under the factors of object deformation,illumination variation and scale variation,this thesis proposes an improved object tracking algorithm based on full convolution Siamese network.The feature extraction network of the algorithm combines the advantages of AlexNet and VGG16 network structure.It not only reduces the number of neural network parameters,but also expands the local sensing field of convolutional layer and improves the representation ability of the model to object appearance features.Group normalization layer and optimized activation functions are added to the network to accelerate the convergence speed of the model.The experimental results show that although the processing speed is slightly slowed down,the improved algorithm is more robust than the benchmark algorithm under the interference of object deformation,illumination variation and scale variation.Aiming at the background clutter and especially similar background,it is difficult for many object tracking algorithms to accurately track the object.This thesis proposes an object tracking algorithm based on multi-feature fusion.The framework consists of Siamese feature extraction subnetwork and region proposal subnetwork.The Siamese feature extraction subnetwork extracts the feature of the sample frame and detection frame respectively,and integrates the extracted deep semantic information and shallow visual information,so as to improve the recognition ability of the algorithm to the object under similar background.Region proposal subnetwork contains classification and regression branches,in which the classification branch calculates the probability of including the object in the candidate region,and the regression branch carries out regression to the location and size of the object boundary box.In the inference phase,the algorithm inputs the feature maps of the sample frame and the detection frame obtained by the Siamese feature extraction subnetwork into the classification branch and regression branch of the region proposal subnetwork,and the feature map of the sample frame is used as the convolution kernel to convolve the feature map of the detection frame.Finally,the algorithm outputs a more compact object bounding box.It is verified by experiments that the tracking performance of the algorithm is better than benchmark algorithm under similar background interference.At the same time,the improved algorithm has real-time performance.In this thesis,the object tracking algorithm based on deep learning is studied.Two more robust object tracking algorithms based on Siamese network are proposed.The experimental results shows that the improved algorithm achieves a good balance between tracking accuracy and real-time performance.