| As a new image acquisition technology emerging in recent years,Unmanned Aerial Vehicle Imagery has been more and more widely used in transportation,photography,security,mapping and other fields.Using road images obtained by UAV and detect road vehicle targets real-time and accurate can provide strong data support for intelligent transportation systems and accelerate the process of smart city construction.However,UAV road images are restricted by factors such as small vehicle size,unbalanced number of vehicle categories,complex and changeable background,etc.It is difficult for current object detection methods to achieve accurate and real-time detection.In this thesis we propose a multi-scale feature fusion vehicle detection algorithm based on deep learning,which achieve fast and accurate small vehicle detection in complex traffic scenarios.The thesis includes the following two contents:Firstly,this thesis analyzes the difficulties of vehicle detection tasks in UAV imagery road scenes,and proposes a multi-scale feature fusion vehicle detection algorithm with AnchorBased detection mechanism.The algorithm sets corresponding anchors according to the size distribution of vehicle data,which improves the matching effect of the network on small vehicle targets.The algorithm also uses a multi-scale adjacent feature fusion architecture,which provides highly semantic features and improves the network’s feature expression ability.In addition,in order to solve the problem of sample imbalance during training,this thesis proposes an alternating training strategy using multiple loss function,which utilizes the advantages of different loss functions to train the network and improve detection performance.Based on the above improvements,this algorithm initially realizes the accurate detection of small vehicles objects from UAV imagery,and can meet the real-time requirements.Secondly,in order to solve the Anchor-Based detector shortcomings such as complicated anchor settings,redundant calculations,missing and wrong matches,this thesis introduces an Anchor-Free detection mechanism in the multi-scale detection algorithm above,making the network no longer reliance on anchors and become simpler and more efficient.At the same time,this thesis uses the advantages of the Anchor-Free detection mechanism to further improve the multi-scale feature fusion architecture,which improves the detection accuracy.Experiments show that the UAV vehicle detection algorithm can effectively identify various categories of vehicles in complex traffic scenarios,its detection accuracy can reach 95.7%m AP on high resolution UAV vehicle dataset,and the detection speed can reach 22 FPS.In summary,the multi-scale feature fusion vehicle detection algorithm proposed in this thesis can fast and accurately identify small vehicle objects from UAV imagery,and the application of the Anchor-Free detection mechanism enables the algorithm to easily and quickly migrate to other detection tasks,which has certain theoretical research value and practical application value. |
