Research On Fast Detection Algorithm Of Vehicle Exhaust Target And Embedded Device Deployment

With the improvement of the national economy,the number of motor vehicles in China is increasing rapidly,and the exhaust emissions from motor vehicles have caused serious air pollution problems.The exhaust pollutants caused by motor vehicles will not only destroy the ecological environment,but also cause great harm to human health.The local governments in China have successively issued a number of air pollution prevention and control documents,and the documents clearly state that it is necessary to strengthen law enforcement inspections of mobile sources such as motor vehicles and punish smoky vehicles according to law.At present,the supervision of smoky vehicles by the environmental protection department is mainly to use smoke meters to sample and inspect the motor vehicles exhaust This supervision method is inefficient,and long-term work will cause great damage to the health of inspectors.Therefore,there is an urgent requirement to develop a high-precision,non-contact fast detection method for smoky vehicles.Using target detection technology to build a motor vehicle exhaust target detection system can meet the requirement,and is very suitable for law enforcement detection of smoky vehicles.Aiming at the requirement of law enforcement detection of smoky vehicles,this paper proposes a fast detection algorithm of vehicle exhaust targets,which is deployed on embedded devices for comparative analysis experiments.Based on the principle of target detection algorithm,this paper compares and analyzes the advantages and disadvantages of the Yolo series algorithm on the self-built vehicle exhaust dataset,and establishes Yolov5s as the basic model for the research.Due to the problems of light transformation,vehicle occlusion and vehicle tail shadow in the actual detection process,this paper uses brightness transformation,saturation transformation,contrast transformation and Cutout data augmentation methods to expand the dataset to 6102 images,which greatly improves the detection accuracy.For the purpose of rapid detection,TensorRT is used to accelerate the network model for NVIDIA devices.In addition,a depthwise separable convolution and attention mechanism are used to improve the Yolov5 network,and a lightweight target detection algorithm Yolo-Light is proposed,which is deployed on an embedded device(Jetson Nano).The main contribution of this paper is the establishment of vehicle exhaust data sets.And in view of the difficulties in the actual detection process of motor vehicle exhaust,brightness transformation,saturation transformation,contrast transformation and Cutout data augmentation methods are adopted to expand the dataset,so that the detection accuracy is greatly improved.Moreover,a lightweight target detection algorithm Yolo-Light is proposed,which can achieve the detection speed of 16FPS on embedded devices,and the recognition rate of smoky vehicles can reach 91.57%.