| Pedestrian intention recognition is of great significance for the intelligent driving of unmanned vehicles and the construction of intelligent transportation system in the future.However,at present,the main research direction in the field of artificial intelligence focuses on pedestrian detection,and there are few methods and Research on pedestrian intention recognition.This dissertation focuses on pedestrian posture and attention mechanism to study pedestrian crossing intention recognition,which is divided into three parts: pedestrian target detection,pedestrian posture estimation and crossing intention recognition.The main contents are as follows:1.A pedestrian target detection method based on improved YOLOv4 is proposed.Aiming at the problem that the detection accuracy of YOLOv4 algorithm is reduced in actual scenes such as unclear light,the way of hole convolution is introduced to strengthen the ability of network feature extraction;At the same time,the anchor frame preset in yolov4 is not completely suitable for pedestrian detection in this dissertation.By analyzing the characteristics of pedestrian aspect ratio,the anchor frame is re obtained by K-means algorithm.Through comparative experiments,the pedestrian detection accuracy of this method in complex scenes is 85.1%,which has better detection effect than the original algorithm.2.A pedestrian pose estimation method based on improved high-resolution network is proposed.Aiming at the problem that the fourth stage of HRNet has little improvement on the detection performance of key points and increases a large number of network parameters and computation,the network structure is appropriately simplified by removing the fourth stage to improve the detection rate.At the same time,in order to ensure the accuracy of the algorithm,U-Net network is introduced as the feature extraction module.Structurally,the U-Net modules are properly adjusted and connected in parallel to the simplified HRNet network to keep the high-resolution characteristics unchanged.Through comparative experiments,the network parameters of this method are reduced by 17.8MB,the average accuracy is improved by 1.1%,and can deal with partial occlusion.3.A GRU pedestrian crossing intention recognition method integrating attention mechanism is proposed.The pedestrian crossing intention recognition is transformed into a time-series prediction problem for modeling.The improved pedestrian detection algorithm is used to obtain the multi pedestrian trajectory information,and the improved attitude estimation algorithm is used to obtain the key point feature combination information.The two are taken as the time-series pedestrian crossing feature information.By constructing a GRU model integrating the attention mechanism,the weight of different feature information is allocated,and finally the pedestrian crossing intention recognition is realized.Through experiments,the accuracy and F1 score of this method are 80.9% and 81.4%respectively,which has better recognition effect.The research content of this dissertation can be applied to the field of unmanned vehicle driving.The crossing intention information provides the basis for unmanned vehicle decision-making.At the same time,in the field of intelligent transportation,it provides the prerequisite for developing pedestrian crossing early warning devices and solving the contradiction between people and vehicles. |
