Pose Estimation Method Research For Excavator Based On Deep Learning

With the vigorous development of China ’s construction industry,the safety problems of construction sites are also showing a multiple trend,and the complex dynamic characteristics of construction machine are important factors of safety accidents.According to the latest data(General Office Of Ministry Of Housing And Urban-Rural Development,2020),689 accidents and 794 deaths occurred in 2020.The security situation is still grim.Among them,the injury accidents caused by construction machines exceeded 30%.This is due to the complex operation,heavy weight,and powerful hazards of construction machines,which often the most prone to latent danger in the construction site.Therefore,monitoring the real-time position and pose of construction machine is of primary importance for the safety management of construction sites.Considering that traditional construction machine safety monitoring and evaluation methods cannot adapt to the complex construction environment,and the monitoring methods based on sensor equipment cost too much.This paper aims to introduce computer vision and deep learning technologies to propose the YOLOv5-Fast Pose(YFP)model to realize the pose estimation of construction machines such as excavators by improving the Alpha Pose human pose model.The model in this paper introduced the object detection module YOLOv5 m to improve the recognition accuracy for detecting construction machines.Meanwhile,to better capture the pose characteristics,the Fast Pose network optimized feature extraction was introduced into the Single-Machine Pose Estimation Module(SMPE)of Alpha Pose.Because there must be corresponding construction machines data support before constructing the construction machines pose estimation model,this paper used Alberta Construction Image Data Set(ACID)and Construction Equipment Poses Data Set(CEPD)to establish the data set of object detection and pose estimation of construction machines.At the same time,for the problem of small sample size of existing data sets,this paper training and testing the YFP model through data augmentation technology and Labelme image annotation software.The sample size is more and more balanced.In this paper,the YOLOv5-FastPose(YFP)model constructed above is verified by the " performance evaluation experiment of the CMOD model " and " the performance evaluation experiment of the YFP model ".The experimental results show that the CMOD model trained in this paper has an Precision of 0.913,a Recall of 0.953,a m AP@0.5 value of 0.973,and an Inference time of 7.3 ms when the confidence threshold is 0.5,and the IOU threshold is 0.5.In addition,the improved model YFP proposed in this paper achieves good results in terms of average Normalization Error(NE),average Percentage of Correct Keypoints(PCK)and average Area Under the PCK Curve(AUC).Compared with existing methods,The YFP model in this paper has higher accuracy in the pose estimation of the construction machine.Method proposed in this paper extends and optimizes the human pose estimation model Alpha Pose to make it suitable for construction machines,improving the performance of pose estimation for construction machines.This paper not only expands the application scope of computer vision technology,but also promotes the technological progress of intelligent construction sites.Its practical significance is to help improve the automation of construction site management,and accurately grasp and implement the strategic decisions of the national and Sichuan provincial governments.It will provide strong support for the high-quality development of intelligent construction and better promote national development and social progress.