Research On Real-time Vehicle Tracking Algorithm For UAV Aerial Video

Vehicle tracking is one of the research hotspots in the field of computer vision and is widely used in intelligent transportation system.Using UAV aerial video to track vehicles can increase the monitoring range of the traffic monitoring system,and solve the problem of the limited monitoring range of the traffic monitoring system caused by the traditional fixed camera.Therefore,it is of great significance to study the accurate and stable tracking of vehicles in the drone scene.However,there are difficult challenges in traffic monitoring of the drone scene.Viewing angle change,rapid vehicle movement,more similar vehicle interference,camera movement and other issues make the tracking effect of commonly used algorithms unsatisfactory.Therefore,this thesis combines the characteristics of the drone scene and analyzes the shortcomings of SiamFC algorithm to improve the tracking performance.Two aspects of work are done in this thesis.(1)A siamese network tracking algorithm based on hard sample mining is proposed.The environment of UAV aerial video is more complex than that of fixed camera.Only the last layer features is used in SiamFC algorithm to characterize the target.When the appearance of the target changed,the tracking robustness is very poor.And in the training process,the imbalance of positive and negative samples and a large number of simple negative samples are not considered.The network cannot learn discriminative information well.Therefore,it is easy to track failed when similar background interference occured.First of all,a feature fusion module is designed to fuse the features of the last three layers extracted by the backbone network to obtain more robust features.Then a hard sample mining loss function is proposed to strengthen the learning ability of network and the discriminative ability of tracking model.The algorithm in this thesis is tested on the OTB2015 dataset and GOT10k dataset.The result of OTB2015 show that this algorithm increases the success rate by 2.6%and the accuracy by 2%compared with SiamFC.On the GOT10k dataset,the mAO of this algorithm is 3.7%higher than SiamFC.The test results verify the effectiveness of the proposed algorithm in this thesis.(2)A vehicle tracking algorithm based on deep siamese network is proposed.In the drone scene,there are many problems such as vehicle deformation,similar vehicle interference,rapid vehicle movement and camera movement.However,the backbone network of SiamFC algorithm is a relatively shallow AlexNet.The features ectracted by AlexNet lack semantic information,which can easily lead to tracking drift or even tracking failure.In addition,due to the limitation of power consumption,complex algorithms are difficult to achieve real-time tracking.First of all,according to the lightweight thinking of GhostNet,a magic residual cropping module and a downsampling magic residual cropping module are designed.And then these two modules are used to construct a lightweight deep network to extract features which include rich semantic information.This lightweight deep network avoids a large increase in parameters and calculations.While ensuring real-time performance,it improves the robustness of vehicle tracking in the drone scene.In addition,an adaptive global cross correlation operation is designed to integrate the global thinking of Non-Local into the cross correlation.The global information of the target is considered when calculating the similarity,which further enhances the discriminative and adaptive capabilities of the tracking model.This algorithm is tested on the UAV123 dataset and UAVDT dataset.On the UAV123 dataset,this algorithm increases the success rate by 5.1%and the accuracy by 3.4%compared with SiamFC.On the UAVDT dataset,the success rate of this algorithm is 1.1%higher than SiamFC,and the accuracy of this altorithm is 4.1%higher than SiamFC.The test results show the effectiveness of the proposed algorithm for vehicle tracking in drone scene.