| HOV(High-Occupancy Vehicle)lane is a multi-occupant lane,a dedicated lane that is only open to multi-occupants within a specified period.By setting up HOV lanes,existing road resources can be effectively used and transportation efficiency can be improved and the role of alleviating traffic congestion has also been achieved,but how the HOV lanes are accurately enforced has always been a problem that needs to be solved urgently.The low efficiency of law enforcement in domestic HOV lanes and the poor practicability of law enforcement methods based on infrared thermal imaging detection methods have been considered,in this thesis,the interior image of the vehicle cabin obtained by the multispectral HOV lane dedicated imaging device is used as the basis for the research on the number of vehicle occupants.The image is similar in visual features to infrared images.The occupant target is overexposed and the features of the occupant targets are pretty different,due to the influence of the weather,light and other factors.Therefore,the improved deep learning related algorithm in this thesis is adopted,and the detection of the number of vehicle occupants is achieved.The main contents of this thesis are as follows:(1)In order to improve the detection accuracy and detection efficiency of the number of occupants,the detection data needs to be pre-processed in advance.First,the Mobilenet V3 lightweight classification network is used to screen law enforcement vehicles,then the Lite RASPP semantic segmentation algorithm is used to extract the detection area of the vehicle cab.Finally,aiming at the overexposure problem of the occupant target,the Pix2 Pix GAN network is used to enhance the image data to fit the ideal image.At the same time,the data generated after the image enhancement is added to the training set to enrich the training samples.(2)The Faster R-CNN algorithm is used as the basis.After the detection accuracy,speed and other indicators are comprehensively considered,Mobilenet V3-Large is selected as the feature extraction network,and a feature enhancement module based on Ghost Module is proposed.ROI-Align is used instead of ROI-Pooling to improve the feature quality after feature mapping in the candidate region.The K-means clustering algorithm is used to obtain the priori distribution of the length and width of the detection target from a large amount of training data,and the rationality of the candidate frame generated by the RPN network and the speed of position regression are improved.In the case of multiple occupants,KL loss is introduced for problems such as false detections and missed detections caused by occlusion between occupants.At the same time,the combination of Soft-NMS and variance voting is used to improve the rationality of the NMS filtering process of repeating target frames.The stability of position regression and the predictive ability of overlapping are improved.(3)The training and detection plan of the detection algorithm is determined through experimental analysis,and then based on NVIDIA’s Jetson Xavier NX platform,the detection algorithm in this thesis is developed and optimized,and the inference process is accelerated through Tensor RT technology,the real-time performance of the detection algorithm,Marginal computing power and practical value have also been improved.The above algorithm improvement content in this thesis has been developed and tested on data sets in actual application scenarios.The experimental results show that the execution efficiency of the overall detection algorithm is improved after the detection image is preprocessed,and the missed detection rate is within the acceptable range.After the algorithm is improved,the task of detecting the number of vehicle occupants in the presence of occlusion and unsatisfactory imaging effects can be well completed.The requirements of the detection accuracy of various types of occupants in the industry that are greater than 80% can be met.The marginal deployment of detection algorithms has also been achieved. |
PDF Full Text Request