Research On Smoky Vehicle Detection Technology Based On Computer Vision

As the number of motor vehicles continues to rise,the illegal emission of exhaust gas has posed a significant threat to the environment and the health of residents.In the early stage,the detection of smoky vehicles mainly relied on human resources,which resulted in resources wasting and high rate of omission factor.The rapid emergence of computer vision technology has opened up new development opportunities for smoky vehicle detection in recent years.For the problem of smoky vehicles detection,this dissertation studies the smoky vehicles detection technology based on computer vision by using two research ideas,which are smoky vehicles detection based on location relationship and classification ideas.The main work is as follows.For the current situation of the small dataset of smoky vehicles,the data enhancement method is used to stitch the smoke pictures with similar features of black smoke onto the road background pictures,which is used to assist the training.This method has solved the current problem of small dataset of black smoky vehicles,and the experiment has proved that the mean average precision is improved by 6.7% after using the assisted training.Since the conventional motion target detection algorithm extracts incomplete target contours and it is difficult to distinguish the background from the target under the moving camera device,the YOLOv3 deep learning target detection algorithm is used to detect the vehicle and black smoke simultaneously.To solve the problems of large computation and complex network model,a lightweight network YOLOv3-M3-CBAM is developed and then further optimized.Experiments show that the improved network has a reduced computation and number of parameters,a detection speed of 21 FPS,and the mean average precision of 93.8%.Based on the target location and tracking information output from YOLOv3-M3-CBAM,smoky vehicles detection algorithm based on the location relationship is studied to determine the number of smoky vehicles in the video.The experimental results show that the recall of the algorithm is 85.3%,and the precision is80.6%.For the problems of YOLOv4 such as large computation and model size,the GhostNet network is used to replace the YOLOv4 backbone network for vehicle detection.The experiment proves that the improved YOLOv4 reduces the model size and computation.The mean average precision is 96.3%,and the detection speed is 20 FPS.The smoky vehicles detection algorithm based on the classification idea is designed by locating the region of interest at the rear of the vehicles according to the detection results.EfficientNetv2 is used to classify the region of interest and combine the tracking information to determine whether the vehicle is a smoky vehicle,and output the number of smoky vehicles.The experimental results show that the recall of the algorithm is 88.2%,and the precision is 81.9%.