Research On Pedestrian Intention Recognition Technology Based On Deep Learning

There are many researches about pedestrian in the field of artificial intelligence.However,there are a few researches on the recognition of pedestrian crossing intentions,which has great significance in the field of intelligent driving.This paper focuses on the method of pedestrian crossing intent recognition based on deep learning,which is mainly divided into pedestrian detection,pedestrian tracking,and the recognition of pedestrian crossing intention.The main contents are as follows:Firstly,we propose an improved pedestrian detection network framework based on YOLO.Aiming at the problem of high miss rate of pedestrians in detection task,the candidate frames are initialized to make them more consistent with the aspect ratio of pedestrians,and then a deeper network is designed to obtain richer feature semantic information.The network is adjusted using the ResNet network model,which significantly reduces miss rate of pedestrian.In order to further reduce the miss rate for small targets,it is proposed to fuse the semantic information of high-level convolution features with the spatial information of low-level convolution features.And the pedestrian miss rate significantly decreases through multi-scale detection.Secondly,we propose a fully-convolutional network target tracking algorithm based on correlation filters.The correlation filter is an efficient tracking algorithm,but the tracking success rate is not high.To solve this problem,the forward and backward propagation formulas of the correlation filter are derived to realize the design of correlation filter template in a full convolutional network,and complete the end-to-end training framework for pedestrian tracking.Experiments show that the method achieves a higher tracking success rate under scale changes and occlusion conditions.Thirdly,we propose an LSTM pedestrian crossing intention recognition algorithm based on attention mechanism and we consider pedestrian crossing intention recognition as a time series modeling prediction problem.Aiming at the problem of multiple pedestrians crossing intention in video sequences,we combine detectors and trackers and a multi-pedestrian trajectory extraction framework is proposed to perform pose estimation and feature extraction for pedestrians in each trajectory.For the problem of assigning different weights to features,the attention mechanism is incorporated in the LSTM model and key features are assigned more weights.Through experiments,this method has obtained Higher accuracy and F1 score.The research in this paper can be applied to self-driving cars,which can provide the basis for driving decisions of self-driving cars by identifying pedestrian crossing intentions.