Distance Estimation in Physical Test Based on Key Frame Localization and Salient Target Discrimination

Current sports testing relies heavily on expert assessment, which is labor intensive and can introduce human error. In this paper, we propose a novel architecture for video-based distance estimation, which can be realized to estimate the distance information of physical exercise from a single video, thus reducing the occupation of other resources. Our approach consists of three main components: key frame localization, salient target differentiation, and distance estimation. A video comprises numerous frames; thus, pinpointing the key frames pivotal for temporal sequencing in distance measurement tasks is essential. Subsequently, one must identify the skeletal key points within these key frames. Given the presence of multiple individuals and objects within a scene, it is crucial to discern and extract the pertinent skeletal key points of the human subject under observation, as well as those of any relevant instrument. Utilizing this extracted data, one can then proceed with the computation of distance. Experiments in seated forward bends and broad jump demonstrate the effectiveness of our approach where the distance estimation results are within error tolerance and provide a new idea for video distance estimation. This method also provides detailed insights into movement dynamics, offering valuable insights for professionals in the fields of sports science and physical education.