Research On Smoky Vehicle Detection Technology Based On Deep Learning

With the improvement of people’s living standards,people pay more and more attention to environmental protection.The remediation of vehicles with excessive emissions is an important part of it.The intelligent detection system of smoky vehicles can obtain evidence for the vehicles with excessive emissions on the road for a long time without human intervention.It is suitable for large-scale deployment and control of urban roads,national highways and other environments,and has a good practical application prospect.In this thesis,the research on the intelligent detection technology of smoky vehicles mainly includes the following aspects:(1)Research on the vehicle detection algorithms.It is known that the shape characteristics of black smoke are not obvious,and it is easy to be confused with the shadow of street trees and vehicles.In order to reduce false alarms,an improved vehicle detection algorithm is designed for the actual vehicle detection scenes.Firstly,the SSD algorithm with the best performance in the vehicle detection task is selected as the backbone after comparison with the mainstream object detection algorithms.Then,the FPN structure and the bounding box for regression without the prior are introduced to solve the problems of inaccurate detection frame and missing large vehicle target.Besides,the loss function design is improved for the problem that the easy-to-detect samples dominate the direction of network optimization.Finally,since the vehicle detection is the pre-process of black smoke classification,the inference time needs to be shortened.Therefore,a lightweight network structure is adopted to realize vehicle detection and achieve the balance between time complexity and precision of the algorithm.(2)Design of the smoke feature classification method based on convolutional neural networks.In view of the small dataset of the smoke classification,we use the transfer learning method and try out the main classification network structures such as VGG,Inception,Mobile Net V2,NASNet and Dense Net in the experiment,and analyze the corresponding experimental results,where the accuracy is improved by modifying AVP to the SPP layer.Then,aiming at the problem that the rear of the vehicle is large in the image,and the effective part that needs to be paid attention to occupies a small image region,we design an attention module named SAM to focus on the spatial information.It can give higher weight to the classification features of the key regions,thus improving the accuracy of the algorithm.(3)Design of black smoke classifier based on spatiotemporal information.Since the twodimensional convolutional network cannot use the temporal information in the video,the single image classification method still has the problem of missing detection and misdetection.Therefore,we design a multi-frame classification network based on 2D-3D fusion in this chapter.The network uses 2D convolution to extract spatial information and 3D convolution to learn spatiotemporal information,which realizes the extraction and utilization of temporal features.Compared with the single frame method,the 2D-3D fusion network has better accuracy performance and classification effect.(4)Software design and implementation.Based on the above algorithms,we adopt modular design to realize the intelligent detection of smoky car.The main modules are: the video image processing module,which is responsible for the batch reading,temporary storage and saving of video images;The vehicle detection module,which realizes the interface and auxiliary functions of the vehicle detection algorithm;The smoke classification module,which realizes the interface and auxiliary functions of the smoke classification algorithm;The vehicle tracking and multi-frame judgment module,which realizes a simple vehicle tracking algorithm and provides support for the multi-frame algorithm.At the end of the thesis,the software functions and the overall algorithm effect and performance are tested.