Research On Detection Technology Of Vehicles With Black Smoke In Video In Real-time Based On Deep Learning

Vehicle exhaust emission seriously threatens the quality of atmospheric environment and people’s health.In response to the harm caused by exhaust emissions,local governments have responded to the national call and issued relevant documents in recent years to increase the supervision of smoky vehicles that emit exhaust emissions in violation of regulations.The traditional detection of smoky vehicles with illegal exhaust emissions is through manual road inspections,periodic sampling inspections,etc.,lack of real-time road detection methods,resulting in low detection efficiency and insignificant governance effects.In order to improve the detection efficiency,a remote sensing detecting method is introduced.The remote sensing detecting results include exhaust emission monitoring data and vehicle video,of which the vehicle video can intuitively reflect whether the vehicle has black exhaust emissions.According to the content of "Result Judgment" of "Measurement Methods and Technical Requirements for Exhaust Pollutants of Diesel Vehicles in Use(Remote Sensing Detection Method)"(HJ845-2017),exhaust emission monitoring data combined with video as evidence can become the basis for punishment by law enforcement agencies,Therefore,a real-time and reliable smoky vehicle detection system based on vehicle video is particularly important.However,due to the fast driving speed of vehicles and the uncertain concentration and shape of exhaust.Therefore,it is a challenge to correctly detect smoky vehicles that emit illegal exhaust in real-time video.This thesis takes the vehicle video collected at the fixed monitoring point of the road as the object,and uses the methods of computer vision and artificial intelligence.Firstly,it detects the moving vehicle in the real-time video,and then identifies whether the tail of the vehicle emits black exhaust.Finally,it adopts the multi frame fusion method to accurately detect the smoky vehicle that emits illegal exhaust.The main research contents are as follows:(1)Designing a vehicle detection method in real-time video based on YOLOv5.In this thesis,the YOLOv5 s object detection algorithm is selected as the vehicle detection model.Considering the monotonicity of vehicle detection samples,the dataset is augmented with left-right flipping,Mosaic augmentation and Mixup augmentation,and data samples are enriched.In addition,in order to speed up the detection speed of the model,this thesis reduces the width and depth of the original YOLOv5 s model,and uses the depth-wise separable convolution operation,realizing the lightweight of the model,and a lightweight detection model YOLOv5 ss is proposed.At the same time,the influence of the channel attention mechanism on the detection performance of YOLOv5 ss is analyzed.The experimental results show that the placement of the channel attention mechanism in the model is different,and the impact on the model is also different.When the channel attention mechanism is used at the end of the feature extraction network of YOLOv5 ss,the detection accuracy of the model is improved to a certain extent.The detection accuracy of the lightweight model YOLOv5 ss with channel attention mechanism is reduced by 0.2% compared with the model YOLOv5 s,but the detection time on a single image is shortened by about 1.5ms.(2)Establishing a black exhaust classification method based on spatial attention mechanism.In this thesis,the Res Net34 classification model is used to classify the image of the rear area of the vehicle to determine whether there is black exhaust.In view of the influence of exhaust gas occlusion and tree shadows on exhaust gas classification of exhaust gas samples,four methods of brightness transformation,saturation transformation,contrast transformation and random erasure are used to enhance the data set,which improves the classification accuracy of the model.In addition,this thesis designs an Efficient Spatial Attention(ESA)module,which is used in Res Net34 to form an improved exhaust classification model ESA-2m Net.The experimental results show that ESA-2m Net does not need to increase the additional classification time,and the classification accuracy is improved by 0.87% compared with the Res Net34 classification model,and the false positive rate is reduced by 3.75%.(3)A multi-frame image fusion method combined with exhaust gas classification confidence is proposed.In order to avoid the contingency problem caused by the final conclusion only based on the exhaust classification results of a single frame,this thesis studies the multi-frame image fusion method of exhaust classification confidence.Based on the position of the vehicle in the video image,this thesis locates the rear area of the vehicle and extracts the pixels of the area,and uses the exhaust classification model to classify this area to identify whether there is black exhaust.In the fusion method,firstly,the tail area image is recognized as the exhaust category,and the picture category with category confidence between 0.5 ～ 0.65 are classified as the non-exhaust vehicle.Secondly,the classification results of all frames in the video are saved.Finally,combined with the classification results of 9 consecutive frames in the video,judge whether the vehicle is a smoky vehicle.If there are 9 consecutive frames of images that are identified as the exhaust gas category,the vehicle belongs to a smoky vehicle;otherwise,it is judged that it does not belong to a smoky vehicle.Compared with the fusion scheme without considering the confidence of exhaust classification,the multi-frame image fusion method combined with the confidence of exhaust classification proposed in this thesis has higher detection accuracy of black smoke vehicles.The detection technology of smoky vehicles in real-time video studied in this thesis is suitable for vehicle videos collected from fixed monitoring points on the road.517 real-time videos(267 smoky vehicle videos and 250 normal vehicle videos)have been tested.The correct detection rate is 87.3%,the false detection rate is 18.8%,the missed detection rate is 12.7%,and the average detection time is 0.95 s,which has the characteristics of high detection efficiency and strong real-time performance.