Motor Vehicle Black Smoke Detection Based On Deep Learning

The pollutant content in the exhaust gas emitted by smoky motor vehicles seriously exceeds the standard,polluting the air quality of cities and endangering the health of the people.Traditional motor vehicle black smoke detection methods usually rely on manual work,which is highly subjective and cannot be continuously monitored,with a high rate of missed detection;The sensor-based detection method has high equipment installation and maintenance costs and poor feasibility.In recent years,with the rapid development of deep learning,object detection has been successfully applied in many scenarios due to its advantages of high efficiency and fast response,which provides a new idea for the detection of motor vehicle black smoke.This thesis realizes the black smoke detection of motor vehicles based on deep learning technology,and the main completed work is as follows:(1)Aiming at the current situation of insufficient motor vehicle black smoke data sets in the research community,a motor vehicle black smoke data augmentation based on StyleGAN is proposed.The black smoke image was obtained through web crawlers and sparse sampling of existing videos.Based on the StyleGAN data generation model,the feature distribution of the black smoke object is learned,and the black smoke patches with high quality are generated.On this basis,the black smoke patches are embedded into the images using progressive weighted fusion to synthesize new samples.The background multi-sampling of the dataset was performed using Mosaic data augmentation,and finally a total of 24638 images were acquired.The similarity between the black smoke patches generated by the StyleGAN model and the real black smoke object was evaluated and compared,and the mainstream object detection algorithm was selected to conduct data augmentation verification experiments.The results showed that the average accuracy of the algorithm had been improved after data augmentation.(2)Aiming at the problems of missed detection and false detection,a motor vehicle black smoke detection algorithm based on multi-scale feature fusion and attention mechanism is proposed.On the basis of YOLOv5,the multi-scale network model is reconstructed,and a shallow feature map is added for feature fusion and detection,which transmits richer fine-grained information and improves the network’s response ability to small target black smoke.When performing multi-scale feature fusion,the Path Aggregation Network is replaced by Bidirectional Feature Pyramid Network to integrate more black smoke feature information and adjust the contribution of different features in the network.The convolutional block attention module of mixed channel attention and spatial attention is embedded between the detection head and the feature fusion part of the network,which further improves the network’s ability to capture key information.Experiments were carried out on the data set,and the experimental results show that the proposed algorithm can improve the missed detection and false detection problems of motor vehicle black smoke detection,and has certain advantages.(3)A motor vehicle black smoke detection system based on deep learning was designed and implemented.The actual demand of motor vehicle black smoke detection is analyzed,and the system processing flow and black smoke vehicle source confirmation module are designed.Transplanting the black smoke detection system to the Jetson TX2 embedded development board has achieved good detection results.There are 41 figures,8 tables and 71 references in this thesis.