Pedestrian Posture Estimation For Smart Car Active Safety

Active safety for pedestrians has always been a topic of great concern in the field of intelligent driving.The on-board camera acquires mainly long-distance low-resolution images.The existing methods for estimating the head and body posture of long-distance low-resolution images mainly adopt the traditional feature extraction methods,and the detection precision is low.The traditional joint estimation method of head posture and body posture is limited to constructing space and time models.The model is too complicated and has low real-time performance.In this regard,this paper mainly carried out the following research work:Firstly,the deep convolutional neural network(CNN)method is applied to the estimation of human head and body posture.Compared with the features manually designed by traditional methods,the convolutional layer in the deep convolutional neural network can automatically learn the optimal features.The back propagation mechanism in the training process makes the network converge to the global optimal solution.The accuracy of Accuracy1 obtained by this method on the CAVIAR head pose data set is 86%,and the accuracy of Accuracy1 obtained on the TUD body pose data set is 83%.Compared with the traditional algorithm,not only the accuracy is significantly improved,but also real-time.Secondly,the deep convolutional neural method based on multi-task learning(MTLCNN)is applied to the joint estimation of human head and body posture.In view of the shortcomings of the existing methods to separate the head pose estimation from the body pose estimation task,Considering the correlation between head pose estimation task and body pose estimation task,the CNN proposed above is used as the basic network to jointly estimate the human head posture and the body posture.The multi-task learning model proposed by this method has stronger generalization performance.Accuracy1 of MTLCNN model is 1 to 2 percentage points higher than that of CNN model on the data set made in this paper,and it has real-time performance.Thirdly,the deep convolutional neural network method based on dual source input(DS-CNN)is applied to the joint estimation of human head and body posture.Since the head region image is lower in resolution than the body image,the head pose estimation accuracy is often lower than the body pose estimation accuracy.Considering the consistency of the human head posture and the body posture,the human head posture feature information is merged with the human body posture feature information,thereby improving the head posture estimation accuracy.Using this method,Accuracy1 is 3.2 percentage points higher than the CNN model on the data set made in this paper.The head posture information improves the generalization ability of the body posture estimation task to a certain extent,and the body posture estimation accuracy is improved by 1.5 percentage points,and has real-time performance.The human head and body posture estimation method proposed in this paper is expected to be applied in the field of intelligent driving.It is used to judge the direction of pedestrian attention and predict the walking trajectory of pedestrians,and make safety warnings in advance to avoid collision accidents and improve road safety.