Vehicle Object Detection And Color Recognition In Rainy And Foggy Weather

Vehicle object detection and color recognition as a branch of object detection has played a key role in recent years for intelligent traffic management and aiding criminal investigation to solve crimes.In real life,we often encounter inclement weather such as rain,snow,fog,sand and dust,etc.,and the quality of images captured in such weather will be greatly reduced.Most of the current high-performing object detection algorithms are trained using images captured in normal weather,so if these models are used directly to detect,the detection accuracy will be greatly reduced.Using image processing technology to improve the accuracy of vehicle object detection and color recognition in rainy and foggy weather is of great significance and practical application value to promote the development of traffic intelligence and assist in the detection of criminal cases.This thesis will focus on the vehicle object detection and vehicle color recognition algorithm under rain and fog weather,and the main work is as follows:(1)A domain adaptive vehicle object detection algorithm in rain and fog weather is proposed.For the problem of difficult collection and labeling of real datasets in rain and fog weather and the domain differences between source domain(normal weather)and target domain(rain and fog),this thesis embeds three domain classifiers on the basis of Faster RCNN to help the network extract domain invariant features in source and object domains;on the one hand,the local features extracted by the shallow feature extraction network of Faster RCNN are aligned to thereby reducing the low-level feature discrepancy.On the other hand,the global features of the image extracted by the deep feature extraction network are aligned to reduce the influence of the overall background layout of the image;then the instance-level features generated by the candidate boxes extraction branch are aligned to reduce the differences in object appearance and viewpoint,etc.Finally,consistency regularization is added to the image-level and instance-level domain classifiers to promote simultaneous alignment of image-level and instance-level,stabilize the model training process,and optimize the alignment effect.The experimental results on multiple synthetic and real vehicle datasets for rainy and foggy days show that the proposed method outperforms other domain adaptive object detection methods and improves the performance of vehicle object detection and its domain adaptive capability under rainy and foggy weather.(2)A joint semantic vehicle color intelligent detection algorithm under rainy conditions is proposed.To address the problems that vehicle color is not easily affected by viewpoint change under rainy weather conditions,the domain adaptive object detection method is slow due to the complex model and the detection accuracy is not high due to ignoring the potential information in the image recovery process,this thesis embeds an image recovery module in RetinaNet for image rain removal,the feature map of the deraining image and input rain image are cascaded sent into the multi-scale feature fusion network behind to accurately extract the joint semantics of the vehicle color feature map,and finally enters the classification and localization subnet.Since the rain removal and detection modules share the feature extraction layer therefore the algorithm maintains the testing time of RetinaNet,and meanwhile improving its robustness in vehicle color recognition under rainy weather.The test results on the self-built rain image vehicle color datasets and the public real rain image vehicle datasets outperforms the most advanced algorithms for vehicle color recognition,popular object detection algorithms and domain adaptive object detection methods.(3)The vehicle detection and color recognition system in rainy and foggy weather is designed and implemented.In order to make the algorithm apply in the actual scene,a vehicle detection and color recognition system in rainy and foggy weather is designed based on PyCharm and Tkinter,integrating the above two algorithm models.The system can be selective detect according to the needs.