Research On Feature Extraction Method Of Vehicle Smoke

With the rapid increase of motor vehicle ownership,motor vehicle exhaust pollution has become an important role in the air pollution in China.Smoky vehicle exhaust pollution is the main source of motor vehicle exhaust pollution.To reduce vehicle exhaust pollution,it is necessary to recognize smoky vehicles from the traffic flow.The vision-based smoky vehicle recognition methods are the mainstream at present,but these methods are vulnerable to miss detections and false alarms.The core to solve this issue is to research on robust smoke features extraction methods.This article focuses on smoke features,including smoke images generation,traditional smoke features,deep smoke features and smoke density features.The main research contents are summarized as follows:(1)A neural network generating smoke images with controllable smoke concentration are proposed.Sufficient smoke images are the basis of smoke feature extraction research,but there are few smoky vehicle images in our dataset.Although existing image generation models can generate smoke images,they cannot generate smoke images with controllable smoke concentration and generate corresponding pure smoke images,which are necessary for subsequent researches.To solve this issue,we propose a controllable smoke image generation neural network to generate various realistic smoke images with controllable smoke concentration and lay the foundation of data samples for subsequent chapters.The proposed model consists of three modules: image separation module,smoke concentration fine-tuning module and image synthesis module,which trained using a three-stage training method.The proposed model generates smoke images by adding smoke to non-smoke images,fine-tuning smoke concentration of smoke images,replacing smoke concentration of smoke images.The model controls smoke concentration in generated smoke images by changing the smoke concentration latent codes in smoke concentration fine-tuning module,and also generates corresponding pure smoke images.The experimental results show that the proposed model is better than existing image generation models in generating smoke images with controllable smoke concentration and generating corresponding pure smoke images.(2)A traditional smoke feature extraction method based on robust volume local binary count pattern is proposed.Traditional feature extraction is suitable for some applications with low hardware configuration.The volume local binary count pattern features not only characterize spatiotemporal texture information of image sequences,but also have the advantages of simple calculation and fast speed.However,from the perspective to characterize smoke image sequence,there are still some deficiencies in the aspects of features anti-noise capability,features non-redundancy,features completeness and multi-scale information utilization.To solve the above issues,we propose a traditional smoke feature extraction method based on robust volume local binary count pattern for smoky vehicle recognition in low hardware configuration.The proposed method uses weighted local threshold instead of the central pixel threshold and introduces an adjustment parameter to balance the noise information and the neighborhood sampling point pixel information,so as to improve feature anti-noise capability.The method takes smoke image and its inverse image as the same class to reduce intra-class distance and feature dimension.In addition,multi-scale information is extracted using multiple sets of different radii and sampling points,and the local information using complete operation is also obtained.The experimental results show that the recognition accuracy of the proposed traditional smoke features is better than existing traditional features.(3)A deep smoke feature extraction method based on enhanced image sequence and submodel fusion is proposed.Deep features have shown excellent performance in the field of feature extraction.Many spatiotemporal feature models can be used for smoke feature extraction.However,these models only extract features from a single RGB image sequence along a single time axis.To make full use of the image sequence information,we propose a deep smoke features extraction method based on enhanced image sequence and sub-model fusion for smoky vehicles recognition in high hardware configuration.The proposed model achieves enhanced image sequence from the aspects of motion information extraction,three orthogonal planes spatiotemporal information extraction and gradient texture information extraction.This method extracts different types of deep features through multiple independent deep models,explores three different sub model fusion methods to achieve complementary features and obtains robust smoke spatiotemporal features extraction.The method obtains complementary spatial features and motion features by integrating spatial network and temporal network and analyzes the same image sequence from multiple directions based on three orthogonal planes to learn different types of complementary features.In addition,gradient and texture image sequences are added to the deep smoke feature to enhance features robustness.The experimental results show that the proposed deep smoke features can effectively improve the recognition accuracy of smoky vehicle.(4)A smoke density features extraction method based on feature enhanced neural network is proposed.The smoke level is an important clue to evaluate the behavior of smoky vehicles.The main method to estimate smoke level is based on comparing the Ringelmann blackness card with the vehicle smoke region manually on site,which is time-consuming and laborious.The methods based on smoke density extraction avoid the above shortcomings.Most of existing models that can be transplanted for smoke density extraction are based on encoder-decoder architecture,in which the feature utilization is realized by directly copying jump connection,However,this kind of direct replication lacks the full use of information of different levels and scales.To solve this issue,we propose a smoke density features extraction method based on feature enhanced neural network to estimate smoke level.The encoder-decoder architecture is adopted in our model.On the one hand,the atrous spatial pyramid pooling module is used to encode the input image at the initial stage and merges different levels of features without losing location information.On the other hand,a variety of feature-enhanced modules are proposed by overlaying the output feature maps of encoding layers to corresponding decoding layers to make full use of spatial features.The multi-scale feature-enhanced module integrates the information of different scales,and the superposition operation integrates the features of different levels to make full use of smoke information.Some popular image segmentation models are improved as baseline methods for smoke density extraction.The experimental results show that the smoke density estimation accuracy of the proposed model is higher than all the baseline methods.In addition,the smoke level estimation results of our method are consistent with the method of Ringelmann blackness card.