Multitarget Dynamic Action Recognition Model for Construction Workers Based on Continuous Action Features

Abstract To address the current problems of frequent injury accidents among construction workers, imperfect supervision systems, and low adaptability of existing algorithms, a dynamic action recognition model for construction workers based on continuous action features and a differential and distance dynamic time warping (DD-DTW) algorithm is proposed. The dataset used in this study is self-collected from real construction sites, including 165,000 video frames covering three typical unsafe behaviors (Clamber, Lean on, Throw) and three common safe behaviors (Stand, Walk, Squat) under different camera angles, lighting conditions and construction environments. First, the multitarget dynamic action recognition model proposed in this study is built through the following modules: feature extraction, input, training set feature integration, feature matching and output. The DD-DTW method, based on first-order differential, second-order differential, and Euclidean distance, was used to analyze the feature similarity of dynamic actions under the time series. In addition, experiments and analyses of dynamic actions for construction workers were conducted, and the DD-DTW, dynamic time warping (DTW), support vector machine (SVM), key point coordinates (KPC) algorithms, long short-term memory (LSTM) algorithm, and convolutional neural network (CNN) algorithm were compared and analyzed. Finally, the real-time recognition of dynamic actions for construction workers was also conducted. The research results show that the average precision of the dynamic action recognition in this study model is 92.33%, the recall is 95.35%, and the F1_score is 93.82%. Its comprehensive performance is much higher than the DTW algorithm, SVM algorithm, KPC algorithm, LSTM algorithm, and CNN algorithm. The average frames per second (FPS) of the algorithmic in this study is 47. This study model can be used for the real-time recognition of dynamic actions for construction workers and is very suitable and has broad application prospects for their safety supervision needs in the construction environment.