Extraction of sports action boundaries using the DeepLabv3+ algorithm

In current sports action videos, it is easy to observe that the athletes’ postures vary greatly and their body shapes are not very aesthetically pleasing. Moreover, as of now, due to the presence of sports equipment or because many athletes are obstructed, these factors can also be considered as influencing factors for determining the boundaries of the actions. Regarding the traditional segmentation methods, very serious problems such as edge fractures, contour adhesion, and detail loss are particularly prone to occur. Therefore, this study aims to conduct research on the extraction of sports action boundaries using the DeepLabv3+ algorithm. In this study, video frames of gymnastics, weightlifting, diving, and sprinting were selected as samples to collect, and image standardization, data augmentation, and manual annotation were carried out before constructing a sports action boundary segmentation dataset. The results of this study undoubtedly provide a more reliable visual basis for sports action recognition, posture analysis, training evaluation, and intelligent auxiliary judging.