Research On Construction Equipment Action Recognition

In recent years,with the rapid development of computer vision technology and computer hardware,action recognition,as a hot research field in computer vision,has been widely used in video monitoring,virtual reality and other fields.At present,the research of action recognition mainly focuses on human activities.In addition,the action recognition of construction equipment has also begun to get the attention of researchers.Construction industry managers can quickly grasp the information of equipment through construction equipment action recongnition(CEAR),make timely judgments and take countermeasures to achieve the purposes of improving production efficiency,saving costs,conserving energy and reducing emissions.However,in the field of construction equipment action recognition,there are problems such as insufficient datasets and relatively insufficient algorithms.In order to solve these problems,this research made three contributions.First of all,this research developed a dataset that can be used for CEAR.The dataset consists of 2064 video clips,including two construction equipment,five action categories,and it is divided into training set,validation set and test set according to the proportion of 6-2-2.Secondly,this research built a deep learning–based action recognition model that combines CNN with LSTM,where CNN is used to extract image features and LSTM is used to extract temporal features from frame sequences.The model achieved an F-1 score of76.25% on the dataset developed in this study,which is comparable to CNN-DLSTM(F-1 score is 75.25%).Finally,this research selected 3d convolutional networks and two-stream convolutional networks as the representatives of human action recognition(HAR)methods,and tested them on the same dataset.The results show that they also have good performance(F-1 scores are 73.55% and 76.26% respectively).Therefore,this research complements the CEAR datasets,provides a simple CEAR model for testing,and preliminarily proves that the HAR methods have the potential for action recongnition of construction equipment.