Research Of Visual Object Tracking Based On UAV Videos

Unmanned Aerial Vehicle（UAV）technology is developing rapidly and has a wide range of practical applications in policing activities such as emergency response,public security patrol,large-scale event security,traffic management,forensic investigation and so on.Driven by high performance computing technology,a batch of object tracking algorithms with excellent performance have been put forward one after another.The research on the object tracking algorithm for UAV videos,which can judge the target position and movement track in UAV aerial images in real time,has broad application prospects.However,the tracking task in this scenario is faced with the challenges of limited computing resources,illumination variations,overcoming small targets,scale variation,low resolution,and background clutter.Therefore,the object tracking method based on UAV video has certain research value.This paper focuses on the lightweight object tracking method based on UAV videos,and the specific work is as follows:In the aspect of lightweight model design,an improved method of lightweight model was proposed to solve the problems of UAV platform’s limited computing resources,inability to deploy large neural networks and real-time tracking task requirements.Firstly,Mobile Net V2 lightweight network was selected to replace Res Net50 network in Saim RPN ++ for feature extraction.Secondly,the improved tracking model is trained from scratch.Experiment results show that the accuracy and success rate on UAV123 dataset are reduced by 0.1% and 1.1%respectively,and the tracking speed is increased by 30.8%,reaching 85 FPS,with a small loss of accuracy for a significant improvement of tracking speed.In the aspect of object tracking task,aiming at the problem that the tracking accuracy of the improved method of lightweight model decreases and tracking drift is easy to occur in complex scenes,an object tracking method combining attention mechanism and triplet loss function is proposed.Firstly,DSA attention mechanism is introduced into the feature extraction stage of lightweight model.Secondly,a triplet loss function is introduced into the classification branch of RPN.Finally,the improved model is trained from scratch.Experiment results show that the accuracy and success rate on OTB100 dataset are improved by 6.7% and 7.3% respectively,and the tracking speed is up to 60 FPS.The accuracy and success rate on UAV123 dataset are increased by 3% and 5.3% respectively,and the tracking speed reaches 55 FPS,which meets the real-time requirements.In software design and implementation,Python language and Py Torch deep learning framework are used to train and implement object tracking algorithm.Matlab GUI was used for interface design in Matlab R2018 b.The software integrated lightweight comparison experiment and object tracking experiment,and software functions were verified on OTB and UAV123 dataset.