Vision-based Algorithm For Pedestrian Motion Perception In Traffic Scenes Research

With the implementation of China’s policy to encourage the development of smart driving cars and smart connected cars,domestic autonomous driving technology has developed rapidly in recent years,but the safety of pedestrians in traffic scenarios has also increased.Therefore,there is a need to enhance vehicle sensing technology to alert drivers to adopt correct driving behavior or to intelligently control vehicles in dangerous traffic scenarios to reduce road traffic risks and safeguard pedestrians.The main research work of this paper is as follows:(1)To judge the scene in front of a car,this paper proposes a Yolov4 target detection model with the introduction of a multi-scale module,which increases the perceptual field of the output feature map of the enhanced feature extraction network,and experimentally demonstrates that the improved model improves the detection accuracy of pedestrian crosswalk and pedestrian at different scales,with an accuracy rate of 0.95 for pedestrian detection and 0.94 for pedestrian crosswalk detection.In addition,we propose the Fast-SCNN semantic segmentation model,which enhances the processing capability of the global feature extraction module for long-range and multi-level feature dependence in images.The method also calculates the overlap between the pedestrian foot area and the pedestrian crossing line to determine whether the pedestrian is within the pedestrian crossing line area.(2)To record pedestrian motion coordinates,this paper proposes a pedestrian tracking and prediction method that introduces fused attention.We add fused attention to the DLA34 of the Center Track model to enhance the extraction of key features of the tracking object and improve the efficiency of information transfer in the network,represent the extracted attributes of the tracking object as points on the heat map,and associate them with a greedy matching strategy to accurately track the object,and then use the LSTM algorithm to loop through the historical positions of the tracking object to predict its future position.Finally,the predicted positions are visualized and displayed.The experiments show that the improved Center Track model achieves a MOTA of 91.9,MOTP of 80.6,MAE of up to 6.13 px,and RMSE of up to 8.26 px in 1280×720 video,and the visualization results demonstrate that this method can effectively accomplish the task of tracking and predicting the location of pedestrians in road conditions.(3)In order to judge the relationship between the vehicle and the pedestrians in front of the vehicle,this paper uses the Zed2 i depth camera to triangulate the distance of the pedestrians in front of the vehicle.The experiments prove that the absolute error of Zed2i’s distance measurement of the 40 m target in front of the vehicle is 3.69 m and the relative error is 9.23%,which can effectively complete the distance measurement task.The ranging algorithm is also colluded with the Center Track tracking model to represent the pedestrians in the traffic scene with an eye model,and the distance of the vehicle from the eye model in front of the vehicle is compared with the current horizontal collision range and braking distance of the vehicle to determine whether the vehicle and pedestrians are in danger of collision and make corresponding reminders.