Sequence Similarity Measurement for Multi-Human Motion Ability Assessment

There has been considerable advancement in human motion ability assessment utilizing computer vision-based approaches. However, the existing research suffer from ineffective assessment methods and inadequate evaluation indicators. For the multi-person target detection task, a human detection algorithm based on improved YOLOv7 is proposed. The TFC structure is designed to pay attention to the important information in the visual picture and solve the problems of small human foreground area and limb occlusion. The MP-SPD module was introduced to address the detection difficulty of low pixel faces with side face targets. For the task of sequence similarity metrics in the process of motor ability assessment, Hidden Markov Model (HMM) is introduced to improve the accuracy and speed of the similarity metrics of limb movement data. In the public dataset MPII, the improved human detection algorithm can improve the mean average precision(mAP50) by 2.98% compared with the original one while the real-time detection frame rate is guaranteed. Experimental results with 20 subjects show that the proposed HMM-based algorithm for motion sequence similarity assessment has millisecond reasoning speed, good generalization ability and robustness. It has obvious advantages in the performance and efficiency of the multi-person motion assessment applications.