Three-Dimensional Working Pose Estimation in Industrial Scenarios With Monocular Camera

摘要

Three-dimensional (3-D) pose data has drawn great attention owing to its wide range of applications. Internet of Things (IoT)-based techniques have been introduced to collect 3-D pose data. Though previous studies have yielded significant results, researchers have yet to use 3-D pose estimation in real-life applications. Since wearable sensors might be intrusive and infrared depth cameras are sensitive to sunlight, monocular-camera-based computer vision algorithms provide a possible solution. Previous algorithms are trained and tested with simple daily postures. There are industrial scenarios where the poses are more complex and irregular. An example is the poses of workers on construction sites, such as lifting, climbing, and rebar tying. These postures differ drastically from daily postures and vary from person to person. For instance, some workers prefer bending rebar tying, while others prefer squatting rebar tying. As a result, the previous monocular-camera-based-3-D poses estimation methods have proved to be inapplicable to industrial scenarios. Thus, this article developed a monocular-camera-based 3-D estimation method which is suitable for industry working poses. A residual artificial neural network (RANN) with flexible complexity and weighted training loss was designed. A 3-D pose data set, which consists of diversified working poses in worksites, was built to test the performance of the network in complex scenarios. Compared with previous 3-D pose capture methods, the mean per joint position error was reduced by 31.42%. The latency was 0.24 s. Thus, we conclude that the proposed monocular-camera-based method has great potential in industrial application scenarios.