Research On Night Driving Vehicle Detection Based On Information Fusion

With the increase of China’s car ownership and the increasingly severe traffic pressure,night driving safety has become the focus of widespread social concern and research.Effective detection of vehicles driving at night and obtaining comprehensive front vehicle information through sensors,so as to achieve early warning of car collision and intelligent long-range light irradiation,are the basic prerequisites for ensuring night driving safety.However,different sensors have their own performance characteristics,and the use of a single type of sensor cannot fully meet the demand for environmental perception.The method based on information fusion can realize the complementary advantages among sensors and significantly improve the accuracy and reliability of detection.Therefore,an algorithm based on information fusion for night driving vehicle detection is proposed,using millimeter wave radar and vision sensors to detect vehicles ahead at night,respectively.And based on this,a target matching algorithm is designed to correlate the detection results of a single sensor to provide comprehensive and reliable information for nighttime driving safety.The main research of this thesis is as follows.First,design the effective vehicle target priming algorithm based on millimeter wave radar.Based on the completion of the millimeter wave radar selection work and data parsing,the millimeter wave radar-based effective vehicle target priming algorithm is proposed for the problem that millimeter wave radar may output unreliable targets.Experiments prove that the algorithm can fully filter out irrelevant targets and achieve effective vehicle target priming.Next,the vision-based target detection algorithm is designed.From the perspective that vehicles in different states are affected by high beam irradiation and have different danger levels,a nighttime driving road dataset containing 12 target classes is constructed for 139 consecutive scenes.For the consideration of real-time,YOLOv5 l is selected as the benchmark target detection network.To further improve the performance of target detection,an intelligent data resampling method is used to enhance the training data and improve the category imbalance problem of the self-built dataset;and a shared convolution is used to improve the head of YOLOv5 to solve the problem that some of its heads cannot be fully trained due to the unbalanced data scale;meanwhile,to solve the problem that the target labels and The lossless mosaic data enhancement algorithm is proposed to solve the problem of inconsistency between target labels and target features caused by the original mosaic data enhancement algorithm.By combining the above methods,the improved YOLOv5 improves the m AP@0.5 on the validation set by 5.7%,which effectively improves the vision-based vehicle detection.Then,the information fusion-based vehicle far-light shield angle solution method is designed.By defining the far-light shield angle,the non-far-light irradiation area is determined.Meanwhile,the spatial unification of millimeter-wave radar and vision sensor is realized according to the coordinate system conversion relationship,and the temporal unification is achieved by using the least common multiple sampling period.Then the far-light shielding angle solution methods based on millimeter-wave radar and vision information are designed respectively,and the target matching algorithm is developed according to the position relationship between the projection points of millimeter-wave radar on the pixel plane and the target obtained from vision detection,and the detection results of different sensors are correlated to realize the far-light shielding angle solution based on information fusion,which gives more comprehensive status information to the target vehicle ahead.Finally,the information fusion-based night driving vehicle detection algorithm is designed based on the above,and the millimeter wave radar,vision sensors and computing platform are selected to build a night driving vehicle detection system integrating millimeter wave radar and vision sensors.The proposed information fusion-based night driving vehicle detection algorithm is verified using real road traffic scenarios.The experimental results prove that the detection algorithm designed in this thesis can not only effectively obtain comprehensive status information of the vehicle driving ahead in the nighttime traffic scenario,but also effectively reduce the missed detection phenomenon generated by using millimeter wave radar detection alone and the false detection phenomenon generated by using vision detection alone.