Research On Human Keypoints Detection Based On Stacked Hourglass Networks

Distracted driving is one of the main causes of traffic accidents.Distracted driving recognition based on computer vision has the advantages of non-intrusion and low cost,which is the trend of related research.This paper proposes to use the human keypoints detection technology based on computer vision to obtain the keypoints of the driver's upper body,and keypoints helps the network to recognize the driver's behavior by learning the posture information.However,the existing human keypoints detection algorithm has poor accuracy on targets with complex background,severe occlusion,and too small pixels.This paper solves this problem by improving the existing algorithms.The main research contents of this paper are as follows:(1)This paper improves the Faster R-CNN algorithm.Firstly,a character detection dataset called WiderPerson is made in this paper.WiderPerson is superior to existing pedestrian detection datasets in shooting scenes,sample density and occlusion scenes.Secondly,this paper improves the network structure of Faster R-CNN,the improved method includes improving anchors size,using finer feature maps,enhancing RoI feature,setting the ignore sampling region and using dynamic sample strategy.(2)An improved algorithm for human keypoints detection is proposed.This algorithm adds a spatial transformer networks module based on Stacked Hourglass networks,which solves the problem of location sensitivity of Stacked Hourglass networks to character detection results.At the same time,this paper uses the mean square error loss function with hard mining to improve the performance of the algorithm.(3)This paper proposes a distracted driving recognition algorithm based on human keypoints.The keypoints' locations of the driver's upper body are integrated into the classification network,so that the network can identify whether the driver has distracted behavior through the posture of the driver's upper body.The research results of this paper have greatly improved the accuracy and robustness of the human keypoints detection algorithm and the distraction driving recognition algorithm.The accuracy of the final distraction driving recognition model reached 96.934%,which satisfies the actual use requirements.