The field of behavior recognition in construction scenarios. This application field can be technically divided into two parts: video tracking and behavior recognition. After intensive research on the literature this week, I found that it was a big hole. Among them, the most recent papers on video tracking are single-target tracking, and we have to solve multi-target tracking. The most practical one is deepSort; the distance of the camera in the real construction scene, The occlusion of shooting, the uniformity and overlap of workers’ clothing, and the magnitude of illegal actions are all huge challenges; there are many recent papers in behavior recognition, which can be practical and currently finalize the ECO model; in the process of tracking a certain worker’s The extraction of spatio-temporal pipeline data is also a difficult problem and so on. In any case, this hard bone has to be gnawed down.
The training of behavior recognition model is generally to use a camera to collect a video of a certain action for about 2-3 seconds, and then label these small videos with behavioral labels for training; in the test, a certain target must be tracked in combination with the video, and then the target is extracted. In spatiotemporal pipeline data, the background is also important in the extraction of main data.
