Research Of Key Technologies Of 3D Object Recognition On Construction Sites

In the past five years,the construction industry has gradually become the pillar industry of China’s economic development.However,the frequent occurrence of safety accidents on construction sites has become a stumbling block restricting the faster and better development of the construction industry.A construction site is a dangerous place containing many heavy construction vehicles and workers at the same time,and the safety of the workers is always threatened by the construction vehicles.Although,managers will use methods such as arranging security inspectors to conduct inspections and conducting security training to reduce the occurrence of security incidents.However,these guarantees are far from enough.There are still a large number of safety accidents every year,and the safety of workers cannot be effectively guaranteed.While some worksites use technologies such as sensors and radio frequency to warn workers of hazards,they are far from widespread due to factors such as high cost.In order to solve the above problems,the thesis studies the key technologies of 3D object recognition（2D object detection and depth estimation technologies）based on 2D images,and realizes the recognition of workers and construction vehicles in 3D space based on ordinary cameras.The thesis is mainly divided into three sub-tasks:One is to use the 2D object detection method to complete the detection of workers and six types of construction vehicles（excavators,loaders,dump trucks,mixer trucks,road rollers and crane trucks）in the construction site scene;The second is to use the unsupervised monocular depth estimation method to complete the relative depth estimation of the construction site pictures;The third is to complete the scale factor estimation in the construction site scene.The main contributions of the thesis are:1.A dataset of 11,712 images for the detection of workers and six types of construction vehicles at the construction site is created,labeled and divided.And three effective improvements are made to the Cascade R-CNN detection model.And the improved model is transferred to the construction site scene based on this dataset,and achieves the highest mAP of 96.8%.2.Metrics are created to quantitatively evaluate the performance of depth estimation methods.And a few-shot depth estimation dataset for transferring monocular depth estimation models to construction site scenarios in the absence of true depth labels is created.And a labeling method is proposed to label this dataset.3.A new loss function is constructed,and PackNet unsupervised monocular depth estimation model is transferred to the construction site scene based on this loss function and this depth estimation dataset.And based on these evaluation metrics,it achieves an average accuracy of 90.37%and a variance of 0.452.4.A scale factor estimation method is designed for the construction site scene.The feasibility of the method is verified on the Kitti dataset which satisfies all the assumptions made by the method.