Article Preview
Top1. Introduction
Human posture estimation and recognition (Jalal et al. 2020) can be applied to human activity analysis (Murahari et al. 2018; Batool et al. 2019), human-computer interaction (Xu et al. 2019; Grosse-Puppendahl et al. 2017) and visual surveillance (Kumaran et al. 2019; Sharma et al. 2019), which is a hot topic in the field of computer vision. In the process of basketball training and competition, coaches need to make special training plans according to the individual situation of different players. In order to improve the basketball skills of athletes, the traditional training method is that coaches make training plan according to their own training theory and experience by combining with the skill level of basketball players. This training mode is highly subjective and coaches need to spend a lot of time to analyze the athletes' posture. It is difficult to objectively evaluate the athletes' training effect. The core of modern sports training is accuracy and efficiency. If the coach can accurately control the athletes' movement posture training, the effect can be greatly improved. It has become a new research direction to collect and analyze the posture data of basketball players and accurately identify the movement posture by using wireless sensor network (WSN) (Tomić et al. 2017; Wang et al. 2017) and computer vision, which is of great significance to improve the coaches' training plan and the training effect of athletes.
Basketball posture recognition (Fan et al. 2021; Ren et al. 2021) is a kind of human posture recognition. At present, there are two main methods of human posture recognition: posture recognition based on inertial sensor (Wang et al. 2019; Zhang et al. 2018) and posture recognition based on image acquisition (Chevtchenko et al. 2018; Oudah et al. 2020). According to the number of image acquisition devices, posture recognition based on image acquisition can be further classified as monocular video recognition (Cai et al. 2019) and multicular video recognition (Mircoli et al. 2018). The general idea of posture recognition based on image acquisition first adopts the camera to capture the athletes' images or videos, then extracts the hidden motion features, and finally designs a classifier to recognize the athletes' motion posture. The multicular video based posture recognition has high maturity and high accuracy. However, the monocular video based posture recognition needs to be further researched at this stage. First, the human body is composed of multiple components, the motion is very complex, and there lacks 3D information in monocular pictures and videos. The change of human 2D posture cannot be simply described by a unified model. Second, the background and light in the image or video have great changes in different scenes. The clothing of the human body itself do not keep the same as well. These changes result in dramatic variation in the appearance of the human body, which is difficult to be described by a unified model.