Research And Implementation Of Pedestrian Tracking And Pose Estimation In Vehicle Environment

The artificial intelligence technology has been paid more and more attention with the progress of science and technology.In the process of studying artificial intelligence,there are many outstanding scientific research achievements and the deep neural network is one of them.Many researchers try to integrate it into traditional image processing algorithms because of the excellent feature extraction performance of the deep neural network,which makes the computer vision technology has been considerable development.As a result,computer vision technology has been applied more and more widely.The intelligent vehicle is one of the important application directions.Aiming at the problem of low visual perception performance of intelligent vehicles,a pedestrian tracking and pose estimation model based on deep learning,which combines Mask R-CNN and Deep Sort,is proposed.The main contents of this paper are as follows:(1)Aiming at the problem that the anchor box used in Mask R-CNN regional proposal network has a slow convergence speed in training pedestrian target features.The following solution is proposed: according to the statistical law,the anchor box which is more in line with the natural aspect ratio of pedestrian target is used to replace part of the original anchor box.The specific method is to remove the anchor box with the aspect ratio of 2:1 and replace it with the anchor box with the aspect ratio of 2:5.For the new anchor box group,the number and area of the anchor box remain unchanged compared with the original anchor box group,and only the scale of the anchor box is changed.Using this optimization method,the convergence speed of the network,accuracy and speed of detection can be accelerated without introducing extra computation.(2)Aiming at the problem that the direct connection channel is used to connect the input layer and the following layers in the Mask R-CNN deep residual network so that the type of convolution kernel is not enough,the following solution is proposed: transplant the lightweight network module SKNet to replace part of the original network convolution module.The specific method is to use 3 × 3 convolution and 5 × 5 convolution plus a fully connected layer of feature channel weights to form the SKNet network structure.This new network structure is used to replace the convolution module with a 3×3 kernel in the original convolution layer.Using this optimization method,the network can adaptively select the best convolution kernel to improve the quality of feature representation the accuracy and speed of detection during the training process.(3)Aiming to the problem that the traditional convolutional neural network feature is used as the target feature descriptor in the deep sort feature matching process and the expression effect is not ideal after extracting the pedestrian target feature,the following solution is proposed: the directional gradient histogram which is statistically calculated on the grid cells in the local area of the image is used as the new feature descriptor.This optimization method can enhance the anti-interference ability in the process of pedestrian target extraction,and improve the accuracy and robustness of feature matching.The thesis verifies the improvement effect of each key module on the performance of the model through comparative experiments.The proposed model is compared with other mainstream algorithms with the average error distance which between the center point of the pedestrian target output box and the label box.The compared results show that the optimization of the model is practical significance.