Research On Vehicle Detection Method Under Adverse Weather Conditions

Vehicle detection is an important research field of environmental perception of autonomous vehicles.With the development of computer technology and the gradual maturity of sensor hardware,the vehicle detection technology of self-driving cars has reached a high level in good weather conditions,which basically meets the real-time and accurate requirements of vehicle decision-making and control.However,there are still many challenges for vehicle detection in adverse weather environment(night,rainy,snowy,foggy,etc.).Vehicle perception information in adverse weather environment contains a lot of noise,which can easily lead to information loss and resolution reduction of vehicle targets,thus seriously reducing the accuracy and robustness of vehicle detection.The incompleteness and uncertainty of vehicle perception information in adverse weather will directly lead to the failure of automatic driving decision-making planning and motion control,and then lead to traffic accidents.Therefore,how to improve the vehicle detection performance in adverse weather is an urgent and inevitable problem on the road of automatic driving development,and it is also a hot issue in the field of automatic driving.In recent years,scholars at home and abroad have conducted extensive and in-depth research on vehicle detection in adverse weather.After analysis,it is found that there are still the following aspects that need to be improved.First of all,the vehicle detection methods represented by visible light cameras and lidar have inherent shortcomings in adaptability to adverse weather.Although researchers try to improve the vehicle detection performance by means of information denoising and model improvement,these methods are mostly suitable for a single scene and difficult to be compatible with various adverse weather environments.Secondly,millimeter-wave radar detection targets are jumbled,and common effective target recognition methods based on rules and prior knowledge fail to fully consider the influence of weather environment,vehicle motion state and other factors on radar detection information and detection error,thus reducing the recognition accuracy of radar effective targets in complex scenes.Finally,in order to improve the detection accuracy of vehicle targets in infrared images,researchers XII mostly pre-extract vehicle ROI(Region of Interest)by image processing in the image preprocessing stage,but the image-based vehicle ROI extraction method is difficult to eliminate the interference of environmental heat sources,and the extraction accuracy of ROI needs to be improved.In addition,many ROI extraction methods based on images have complicated calculation process,which easily reduces the real-time performance of vehicle detection.In view of the above problems,this paper attempts to explore a vehicle detection method under adverse weather conditions(night time,rainfall ≤25mm/d,snowfall ≤5mm/d,450m≤visibility ≤10km,detection distance ≤60m).Based on millimeter wave radar and infrared camera,advanced artificial intelligence method is used as the technical driver to improve the accuracy,real-time and performance of vehicle detection under adverse weather conditions.The research contents of this paper mainly include the following aspects:Firstly,aiming at the problem of low accuracy of effective target recognition of millimeter wave radar in bad road environment,this paper proposes an effective radar target recognition method considering multi-factor coupling and data driving.Firstly,the influence of weather environment and vehicle motion state on radar detection information and detection error is analyzed,the main influencing factors are identified and coded,and a 14-dimensional multi-factor coupled feature vector is constructed.Then,different machine learning models are selected to train the sample data in a supervised way to generate the basic effective target recognition model.Finally,based on correlation analysis,PCA dimension reduction,feature optimization and grid search,the trained Xgboost model and Light GBM model are optimized.Tests show that the radar effective target recognition accuracy of this model is 97.3%,and the missed detection rate is 2.7%.Compared with the traditional method,the radar effective target recognition accuracy is improved by 3.5%,and the missed detection rate is reduced by 3.9%,which effectively improves the radar effective target recognition performance under adverse weather conditions.Secondly,aiming at the problems of low vehicle detection accuracy and poor scene adaptability in adverse weather.Based on the infrared characteristics of vehicles and advanced artificial intelligence methods,this paper explores three vehicle detection methods based on infrared images.First,to solve the problem of poor adaptability of a single vehicle detection feature scene,this paper proposes a vehicle detection method based on the fusion of edge features and local textures.HOG features representing edge information and LBP features representing local textures are extracted from images respectively,and then HOG-LBP fusion features are generated based on PCA dimension reduction.Finally,SVM vehicle classifier is trained with the fusion features to detect vehicle targets.The test results show that this method can improve the vehicle detection accuracy by 5.6%-8.9% compared with a single vehicle detection feature.Secondly,aiming at the problem that the original Haar-like features can not effectively express the remarkable features of vehicles in infrared images,this paper proposes a vehicle detection model based on improved Haar-like features and Adaboost.This model designs four new Haar-like templates based on the feature distribution of vehicles in infrared images,and extracts features from training samples together with the original Haar-like templates.Finally,Adaboost classifier is trained to detect vehicle targets from infrared images.The test results show that the improved Haar-like features can improve the vehicle detection accuracy by 3.2%.Thirdly,aiming at the problems of large amount of training data and slow convergence in the training process of deep learning model,this paper proposes an infrared image vehicle detection method based on transfer learning and YOLO V4.The test results show that the YOLO V4 model based on transfer learning can improve the vehicle detection accuracy by 1.7% and reduce the model training time by 20%.Finally,three vehicle detection models are tested and compared by real vehicle test.The test results show that the vehicle detection model based on YOLO V4 has the highest vehicle detection accuracy of 92.5% and the fastest vehicle detection speed of 40 Fps.The vehicle detection model based on HOG-LBP+SVM has better detection ability for small target vehicles.The detection accuracy of the improved Haar-like+Adaboost vehicle detection model is similar to that of the HOG-LBP+SVM model,but the real-time performance of vehicle detection has obvious advantages.Thirdly,in order to enrich the perceived information of vehicle targets and further improve the accuracy and speed of vehicle detection under adverse weather conditions,this paper proposes a vehicle detection method based on ROI information guidance and multi-scale information fusion.Firstly,aiming at the problems of poor environmental adaptability and low computational efficiency of image-based ROI extraction method,this paper proposes a vehicle XIV ROI prediction model based on radar information guidance and pixel regression.Experimental results show that the ROI extraction method in this paper has better real-time performance on the premise of ensuring ROI coverage,and the calculation time of a single frame image is only 4 ms,which can save 60%-70% of the calculation time compared with the ROI extraction method based on image.Secondly,the infrared image is reconstructed and enhanced based on the ROI information of the vehicle,which not only effectively reduces the search area of the vehicle target,but also significantly improves the separability of the vehicle target in the infrared image.Thirdly,aiming at the problem that the target fusion method based on two-dimensional spatial information is easily disturbed by adverse weather,this paper proposes a multi-scale information fusion model considering depth information and fusion weight.Firstly,the depth estimation model based on boundary regression successfully expands the fusion information of infrared image from two-dimensional space to three-dimensional space.Then,the weight of fusion information of various scales is optimized by genetic algorithm.Finally,Kalman filter is used to predict and track vehicle targets,which effectively improves the accuracy and robustness of vehicle target matching in complex scenes.Finally,the vehicle detection method under adverse weather conditions is tested and verified by real vehicle test.The content of the test includes scene adaptability,real-time,effectiveness and the accuracy of vehicle target fusion.In the scene adaptability test,the vehicle detection accuracy of each vehicle detection model in different weather is more than 90%,of which the highest detection accuracy is 95.0% at night,and the detection accuracy is slightly lower than 91.5% in rainy days.In addition,the vehicle detection model based on YOLO V4 has the best comprehensive performance in all scene tests.In the real-time test,the fusion of radar and infrared images can improve the vehicle detection speed of 8 Fps—9 Fps,among which the vehicle detection model based on improved Haar—like and Adaboost has the best vehicle detection real-time performance of 42 Fps.In the validity test of the model,the changes of vehicle detection performance before and after information fusion are compared.The test results show that the vehicle detection method after fusion can improve the vehicle detection accuracy by 2.4%-3.7% and reduce the false detection rate by 1.6%-3.0%.Finally,the vehicle target fusion accuracy is tested,and the test results show that the target fusion method in this paper can improve the matching accuracy by 3.2% compared with the fusion method based on two-dimensional spatial information,and effectively improve the target confusion problem in adverse weather conditions.