Research On Real-time Speed Measurement Of UAV Ground Moving Target Based On Learning

The application of communication and electronic technologies in intelligent transportation systems(ITS)has become very common.It has played an important role in monitoring and quickly handling urban traffic congestion and emergencies.UAV are flexible and have a wide field of view.UAV equipped with image acquisition equipment are an important platform for real-time distributed traffic flow supervision.Considering the motion characteristics of drones and the speed requirements for vehicle speed measurement,this article will focus on moving vehicles speed measurement technology based on drone platforms.The main work of the thesis is as follows:(1)Vehicle detection method based on improved YOLOv3 under drone.Aiming at the problem of small target size and unbalanced target ratio of vehicles under UAV,the improved YOLOv3 detection network was proposed for model detection.First,in order to strengthen features' extraction capabilities,a Clique Block feature loop network was introduced to enhance the interaction between local shallow and deep features.Secondly,Focal Loss is used to improve the model loss function,reduce the loss weight of easy-to-classify samples,and increase the model's focus on difficult samples.Then,the GIo U regression loss function is used to solve the sensitivity of the mean square regression loss function to the target size,and the model's attention to small targets is improved.Finally,during the training process,multi-scale training strategy is used to randomly adjust the model input sizes to improve the model's detection performance for targets of different sizes.The final experimental results show that the detection accuracy of the improved detection network on the UAV datasets reaches 95.5%,and the detection rate reaches real-time.(2)Vehicle target tracking method based on data combination.Vehicles are densely on urban roads under UAV,and occasionally there are certain occlusions.This paper studies tracking algorithms based on improved target detection results.First of all,we use Kalman filtering algorithms to predict possible positions of the target,and then extract target motion features and depth appearance features,in combination with the Hungarian matching algorithm,the prediction positions and the detection results of the next frame are matched.Aiming at the problem of tracking trajectory loss caused by target occlusion and missing detection frame,the appearance feature extraction network is optimized,and the cross-entropy loss and metric triplet loss function are used to jointly train the deep network to improve the model's re-identification ability.Experiments show that the algorithm in this paper outperforms better than other algorithms on public data sets,and can achieve continuous and stable tracking of multiple vehicle targets.(3)Method of target speed measurement based on UAV.This solution is based on the target tracking sequence to obtain the pixel displacement of the vehicle target within a fixed time,and then uses the camera calibration method and low-altitude aerial projection principle to convert the pixel displacement into the actual physical displacement to obtain the vehicle target's running speed.Because the actual displacement of the target can be obtained by simple conversion,the algorithm in this paper can efficiently complete the vehicle speed measurement task,and has strong practicability.