Research On Moving Vehicle Detection Algorithm In Real Scene

Black smoke vehicles have a serious impact on the atmosphere.At present,the detection of black smoky vehicles in China is mainly based on manual inspection,but this detection method will consume a lot of manpower and material resources.Automatic detection of smoky vehicles based on road surveillance video can effectively reduce the consumption of human and material resources and has a wide range of application prospects.Some specific types of vehicles are the key targets for black smoky vehicle detection,and filtering these vehicles can effectively reduce the amount of data to be processed and speed up the detection of black smoky vehicles.This thesis studies the algorithm for realistic scenarios of vehicle detection and works on the following aspects.Firstly,the traditional vehicle detection algorithm and the deep learning based vehicle detection algorithm were compared and analysed,and the YOLOv7 detection algorithm was chosen as the base algorithm for vehicle detection.Secondly,to deal with the problems of poor detection of small vehicle targets at a distance,incomplete large vehicle misses and weak detection of obscured vehicles when the basic algorithm detects vehicle models in road surveillance video,an effective feature layer suitable for detecting small targets is added to the original three effective feature layers of YOLOv7;A three-dimensional attention mechanism Sim AM is introduced into the trunk feature extraction network to improve the detection ability of such problems.Finally,the feature fusion module of YOLOv7 is replaced by an ASFF network from the PANet structure in order to make the feature information more reasonable and adequate for fusion,which avoids the limitation that only feature fusion can be performed between adjacent effective feature layers in the original network structure of YOLOv7,and the mutual fusion between effective feature layers is controlled by the weights to enhance the rationality of feature fusion.Also,the position loss function is replaced with SIOU to speed up the convergence of the algorithm.Finally,in order to make the proposed vehicle detection algorithm cope with foggy weather,a generative adversarial network based defogging algorithm is proposed so that the vehicle detection algorithm can still maintain good detection results under foggy conditions.In this thesis,a realistic scenario of vehicle detection dataset is constructed using road black smoke vehicle surveillance video.Comparing the algorithm optimized in this thesis with the original algorithm and the common algorithm for vehicle detection,the detection effect of the algorithm in this thesis is better than other algorithms under the premise that the IOU threshold is set to 0.5 and 0.75,and the detection speed can meet the real-time requirements.