Illegal Vehicle Detection Based On Multi-objective Optimization From UAV Imagery

With the development of our economy and the improvement of people’s living standard,the number of vehicles has been increasing year by year,which cause the phenomenon of illegal parking.Traditional object detection methods of illegal vehicles use manual labeling and traffic cameras,which cost lots of manpower,material resources and have certain constraints.The low size,low resolution and complex background of the illegal vehicles based on the UAV perspective lead to the problem of wrong detection or missing detection in the traditional YOLOv5 target detection network.Therefore,this thesis implements a multiobjective particle swarm optimization of the traditional anchor frame clustering algorithm,while improving the internal modules of the model and adding an attention mechanism to achieve illegal vehicle detection based on muti-objective optimization from UAV imagery.The main work of this thesis is summarized as follows:(1)In order to select the most suitable anchor box of the dataset and solve the problem that traditional YOLOv5 object detection algorithm has weak detection effect on small targets from the UAV perspective,a vehicle violation detection model based on multi-objective optimization and multi-scale perception called MOO-MS-YOLOv5 is proposed.Firstly,the SAMOPSO-Kmeans anchor frame clustering algorithm based on nonlinear self-adapt parameters is introduced to help the speed and position of particles dynamically adjust with the number of iterations,which is able to obtain the most suitable anchor box.Then,three modules respectively called hybrid size cropping(MO-Focus)layer,multi-branch convolutional network fusion(MC-CSP)layer and spatial multi-scale pyramidal pooling(SPN)layer are added in the neck and backbone of the model instead of the traditional modules.Finally,a MOO-MS-YOLOv5 illegal vehicle detection model based on multi-objective optimization and multi-scale perception is constructed.The experimental results show that SAMOPSO-Kmeans can efficiently obtain the most suitable anchor box combination for this dataset and the MOO-MS-YOLOv5 illegal vehicle detection model has a better detection effect on small targets from high altitude perspective.(2)In order to solve the problem that the UAV model is difficult to learn the object features due to the complex visual background,A MOO-MS-YOLOv5 detection algorithm based on LSTM time-sequence cooperative attention mechanism(LSTM-AM)is proposed.Cooperative attention mechanism includes multi-channel attention and temporal attention.Multi-channel attention first sends the location features to the residual module for roll-up operation and global pooling operation,then sends them to the excitation module to obtain the correlation between channels.LSTM temporal attention mechanism is based on a dynamic temporal selection mechanism that computes for global features and motion features.Finally,the ESPCN super resolution reconstruction method is used to improve the clarity of the dataset.The experimental results show that the cooperative attention mechanism algorithm based on multi-channel attention and temporal attention is ablet to ignore the complex background information,make the model focus on the target,and effectively improve the detection accuracy and speed of the model.(3)A UAV perspective vehicle violation management system based on multi-objective optimization of UAV image violation vehicle model is designed and implemented,which uses a front-end and back-end separated system development model.The front-end uses Vue for user interaction page development,the back-end uses Java and Python respectively to realize the specific logic of the system and the core algorithm of the violating vehicle based on multiobjective optimization of the drone image violating vehicle model of the drone view.The system mainly contains four modules,includes user registration and login module,violation vehicle detection module,violation vehicle history query module and violation vehicle data statistics module.These modules help the traffic pipeline department to detect violation vehicles in real time,and to focus on monitoring and rectifying the road sections in areas with more violation vehicles.