Research On Visual Detection Method Of Smoky Diesel Vehicles In Complex Scenarios

Diesel vehicles with excessive emissions emit a large amount of black smoke exhaust during driving,causing serious ecological problems.Therefore,the State Council issued the "Action Plan for the Battle of Diesel Truck Pollution Control" to treat diesel vehicles with excessive emissions."Anhui Province diesel truck pollution prevention and control battle implementation plan" also clearly pointed out to promote the construction of monitoring system,through online monitoring video to achieve all-day emission monitoring.Facing the significant demand of video-based regulation of smoky diesel vehicles,this thesis carries out the research on the visual detection method of smoky diesel vehicles under complex scenes,which is of great significance to assist environmental protection departments in supervision,improve air environment pollution and protect citizens’ health.However,on the one hand,due to the complex and variable scenes where the black smoky diesel vehicles are located,there are many ground background interference factors,the black smoke is not obvious and dissipates quickly,the commonly used smoke detection algorithms are less effective and often not pervasive in static/dynamic complex scenes.On the other hand,urban traffic road scene diesel vehicle emission black smoke process is more complex,there are often difficult to determine its identity through the backward video in the vehicle data,vehicles between each other obscured,across the camera viewpoint vehicle appearance differences and other factors influence are black smoke vehicle identity problems bring greater difficulties.This set of factors poses a challenge for the research of visual detection methods for smoky diesel vehicles in complex scenarios.In response to the above challenges,this thesis carries out the research on the visual detection method of smoky diesel vehicles in complex scenes:1.Aiming at the problem of inconspicuous and fast dissipation of diesel vehicle emission black smoke in complex scenes and the influence of many ground background interference factors,a diesel vehicle black smoke detection method based on motion amplification enhancement and color feature localization is proposed to achieve higher accuracy of black smoke detection in a variety of complex scenes.To alleviate the problem of large data limitations,a diesel vehicle black smoke dataset was constructed,containing diesel vehicle black smoke data for different weather(sunny and rainy),different lighting(noon and evening),and different vehicle models.Design the motion amplification enhancement module to amplify the black smoke features with subtle motion relative to the background considering the spatio-temporal information characteristics.Design the color feature localization module to narrowly focus the black smoke feature extraction area to the key area where black smoke may exist based on the local information of the exhaust tail of the diesel vehicle and remove the interference of the complex background on the ground.Experiments on the diesel vehicle black smoke dataset show that the method achieves 93.61%and 93.56%black smoke detection accuracy in static/dynamic complex scenarios,respectively,with high black smoke detection accuracy.2.In order to solve the problem of identifying smoky vehicles in urban traffic scenarios,we propose a re-localization method for smoky diesel vehicles in cross-camera scenarios,which achieves high accuracy of re-localization of smoky diesel vehicles.To alleviate the problem of large data limitations,a diesel vehicle repositioning benchmark dataset is constructed,containing diesel vehicle repositioning data under different colors,types,viewing angles,resolutions,and lighting conditions.Considering the problem of large differences in vehicle appearance due to different camera views,the IBN module is introduced to build a feature extraction network to improve the adaptability of the network model to changes in image appearance.Design a loss function based on Hausdorff distance metric learning to better measure the difference between two features during network optimization and reduce the effect of occluded samples during the optimization process.Experiments on the diesel vehicle repositioning benchmark dataset show that the relocation effect of our method for smoky diesel vehicles improved by 5%in the Rank-1 index and reached 83.79%in the mAP index with high relocation accuracy compared with the benchmark method.Based on the above research,this thesis realizes the whole technical chain of "black smoke emission detection-vehicle identity identification" for black smoky diesel vehicles in complex scenes,and constructs a visual detection method for black smoky diesel vehicles in complex scenes,which provides reliable data support for the supervision of black smoky diesel vehicles by ecological and environmental departments.