Research On Detection Of Hazardous Chemical Vehicles Base On Deep Learning

Due to the particularity of the cargo carried by the hazardous chemical vehicles,serious losses will be caused once an accident occurs,so it is urgent to supervise the transportation of hazardous chemical vehicles effectively.In recent years,deep learning algorithms have developed rapidly,providing an intelligent theoretical basis for the supervision of hazardous chemical vehicles.Through monitoring equipment,the information of passing vehicles is collected in real-time,and the deep learning algorithm is used for detection,so as to carry out intelligent monitoring of hazardous chemical vehicles.Due to the lack of datasets of hazardous chemical vehicles at present,this thesis firstly obtains the hazardous chemical vehicles dataset by manually marking the vehicles,license plates and dangerous chemical signs in the pictures of hazardous chemical vehicles collected by cameras,and obtains the hazardous chemical vehicles dataset for subsequent model training and testing.Hazardous chemical vehicles usually move fast on the road,and detection systems are usually deployed on mobile embedded devices.In order to ensure the accuracy and real-time of detection,this thesis studies the Mb-YOLOv5 algorithm.Firstly,the SENet(Squeeze-andExcitation Networks)attention mechanism of Mobile Net V3 is changed to 16-fold downsampling CBAM(Convolutional Block Attention Module),and then the original feature extraction backbone of the YOLOv5(You Only Look Once Version 5)network is replaced,which increases the network’s attention to the feature space and reduces the amount of model parameters.Secondly,some of the conventional convolutions are replaced by depthwise separable convolutions with fewer parameters,the size of the model becomes 35.8% of the original,and the detection speed is 40.3% faster.In addition,there is a problem of category imbalance in the hazardous chemical vehicles dataset.For this reason,the original loss function is replaced by the focal loss function in this thesis.On the self-made hazardous chemical vehicle datasets,compared with the previous category with the worst effect,the AP(Average Precision)has increased by 0.13,and the overall m AP(mean Average Precision)increased by 0.026.After the Mb-YOLOv5 algorithm studied in this paper is sparsely trained on the hazardous chemical vehicles dataset,it is still found that the network has a large number of redundant channels.In response to this problem,this thesis further prunes the model,restores the accuracy with knowledge distillation,and then carries out BN(Batch Normalization)layer fusion to further compress the parameters.Through the above operations,The size of the model is only 0.74 M,which is compressed to 15.1% of the original,and the inference speed is improved 33.3%.Finally,after obtaining the target detection results,the LPRNet(License Plate Recognition via Deep Neural Networks)algorithm is used to identify the license plates,and the Deep SORT(Simple Online and Realtime Tracking with a Deep Association Metric)target tracking algorithm is used to obtain accurate vehicle number statistics.In addition,in order to facilitate the management personnel to view and process the monitoring results,this paper implements a monitoring and management system for hazardous chemical vehicles with separated front and back ends based on Java Script and node.js.