Research On Black Smoke Vehicle Detection Method Based On Deep Learning

In recent years,the prevention and control of black smoke vehicle exhaust pollution has become the top priority of environmental protection work.Based on the extensive use of surveillance cameras,it is necessary to develop an intelligent detection system for smoky vehicles based on surveillance video.The main research contents of this paper are as follows:(1)Since there is no publicly available data set of smoky vehicles,9,960 images of smoky vehicles with a resolution of 360x640 were manually collected and marked as the detection data set of smoky vehicles.The data sets of smoky vehicles proposed in this paper are collected in real traffic scenes.(2)The object detection method based on deep learning is studied,which mainly includes two categories: the object detection method based on the box and the target detection method based on the key point.In addition to introducing the classic and representative network architecture of the two types of target detection methods,the basic framework above is also evaluated on MS COCO,the standard data set open in the field of target detection,and the black smoke vehicle data set established in this paper.In view of practical application needs,this paper chooses Center Net as the basic algorithm framework of black smoke vehicle detection in traffic monitoring scene,and conducts research on Center Net.(3)A series of Center Net models based on different backbone networks were tried on the data set of smoky vehicles,among which the Center Net model based on Res Net18 achieved a detection effect of 90.9%m AP on smoky vehicles.Tensor RT was used to accelerate a series of Center Net models based on different backbone networks.The accuracy of all models did not change before and after acceleration and the speed improved significantly.In particular,the Center Net model based on Res Net18 was tested under GTX 1080 and its network inference speed increased to 145 FPS.(4)Aiming at the shortcomings of Center Net model based on existing backbone network in multi-scale detection,this paper designed a new backbone network with multi-scale features,which referred to the multi-scale feature fusion method in FPN and PANet,and combined with Center Net framework,raised the m AP of black smoke vehicle detection to92.68%.With Tensor RT acceleration,the network inference speed can reach 121FPS(GTX1080).(5)When the backbone network in the Center Net framework changed from simple Res Net18 to complex Res Net101 or Hourglass,its detection effect in the smoky vehicle data set did not improve.For black smoke vehicle data set is too small,would not show the complex backbone network said capacity of defects,this paper proposes a based on attention mechanism and combination network backbone network of black smoke vehicle detection,the main improvement is to use two backbone networks respectively extracted black smoke,the vehicle characteristics,including vehicle corresponding backbone network using mass of vehicle detection do preliminary training data set,and design the module of feature fusion based on attention mechanism model on the relationship between the black smoke and vehicles.Based on the above improvements,its corresponding Center Net model improved the m AP of black smoke vehicle detection to 95.19%.With Tensor RT acceleration,the network inference speed can reach 81FPS(GTX 1080).