Research On Target Tracking Algorithm Based On Deep Learning

Object tracking is an important research topic in the field of computer vision and has great application value.The main task is to track the object in subsequent frames of the video when the initial position of the object is given in the first frame.In recent years,the object tracking technology has made great progress,and has been widely used in the fields of video surveillance,human-computer interaction,and transportation.However,due to the complexity of real-world scenarios,there are many challenges in object tracking,such as changes in lighting,complex backgrounds,occlusions,and changes in object scale,and so on.It is a problem for us to design algorithms that meet the real-time and accuracy requirements under complex conditions.At present,deep learning is developing rapidly,and applying deep neural networks to object tracking to improve tracking performance is the main research trend.The main tasks as follows:(1)A correlation filter tracking algorithm based on multi-layer depth features is proposed.In order to solve the problem that the shallow features are limited for object description in the face of fast object motion,complex background and occlusion.Under the correlation filter tracking framework,our algorithm uses the VGG-19 convolutional neural network to extract the depth features of different convolutional layers and fuse the response graph to achieve object tracking.At the same time,the feature selection algorithm is used to reduce the feature dimension and improve the tracking speed.The experimental results on the OTB2013 dataset show that the accuracy rate is 88% and the success rate is 87%.(2)A siamese network object tracking algorithm based on re-detection mechanism is proposed.To solve the problem of tracking failure of SiamFC algorithm when encountering situations such as occlusion,similar objects,and complex backgrounds.Under the framework of deep learning tracking,a detection module is introduced,during the tracking process,the threshold is used to determine whether object re-detection is required.At the same time,a high-confidence model update strategy is adopted to ensure that the model is updated reasonably and to ensure the tracking speed.The experimental results on the OTB2013 dataset show that the accuracy rate is 88.4% and the success rate is 85.7%.