| In the road traffic scene,the safety of pedestrians,as road users,not only affects the traffic order,but also affects the road risk state.In order to improve the safety of pedestrians and reduce the occurrence of traffic accidents,it is necessary to judge whether pedestrians cross the street.The traditional pedestrian intention discrimination methods mainly rely on a single sensor to detect,and judge and predict by detecting the pedestrian’s limb action or trajectory route.However,because the pedestrian motion belongs to nonrigid motion,the motion is flexible,and these methods ignore the influence of road scene.Based on the on-board mobile equipment,this paper integrates the detection results of lidar and camera at the decision level,analyzes the influencing factors affecting pedestrian crossing,and judges the pedestrian crossing intention.It is mainly divided into the following three parts for research,which are sensor fusion method,pedestrian detection and tracking and pedestrian intention recognition.（1）In this paper,the two sensors are calibrated in time and space and the distortion is corrected to process the original data,analyze the available data in the traffic scene and different types of fusion methods,and choose the fusion method reasonably.At the decision-level fusion level,using the basic theory of D-S evidence theory,according to the distance between samples,the weighted decision fusion and the area ratio of the detected overlapping parts are used to judge the category and consistency of the target.Due to different sensors and different types of information obtained,compared with a single sensor,its information source is more accurate and complete.The fusion algorithm determines the final result after target detection.（2）The detection of lidar first further processes the corrected original data,analyzes the characteristics of lidar data,calculates the effective range of the target that can be detected by lidar,and performs threshold segmentation according to this range index to reduce the complexity of the data.In terms of algorithm processing time and accuracy,DBSCAN,two improved DBSCAN clustering algorithms and the complex YOLO algorithm based on deep learning are compared and analyzed to improve the efficiency of lidar detection.The detection of this part of the video image mainly analyzes the network structure of YOLOv3 and YOLOv4,and selects four evaluation indicators for comparative analysis.The Sort algorithm is used to track the target,and the improved algorithm Deep Sort is used to solve the problem that the target number is confused due to occlusion.（3）According to the result of detection and fusion,by analyzing the influence of environmental factors and pedestrian factors on pedestrian crossing,the Bayesian model is used to discriminate pedestrian crossing intention.Count whether pedestrians have interactive behaviors,whether pedestrians observe vehicles,whether pedestrians are safe to cross the street,whether pedestrians wait on the side of the road,and whether traffic facilities allow pedestrians to pass through these five factors to analyze the impact on pedestrians crossing the street,and give each factor’s impact on the intention of crossing the street.degree of influence.The pedestrian intent recognition method based on radar video fusion proposed in this paper is helpful for the study of pedestrian crossings by vehicles in intelligent driving scenarios,and provides a basis for subsequent deeper research. |
