Intelligent Assessment of Healthy Behavior in Digital Exercise Development Based on Pose Estimation and Deep Learning

To improve the intelligence and scientific rigor of sports behavior assessment in digital sports scenarios, this study proposed a deep learning–based motion behavior recognition and intelligent evaluation framework that integrated “recognition + evaluation” into a unified system. First, a pose estimation model was employed to extract skeletal keypoints from sports videos with high precision. Long Short-Term Memory (LSTM) and Transformer architectures were then combined to model temporal pose sequences and enable multi-class sports behavior recognition. Existing research primarily focused on isolated recognition tasks and lacked finegrained quantitative evaluation of movement quality. Moreover, many systems depended on standardized experimental environments, resulting in limited generalization ability and weak support for personalized feedback. To address these limitations, a prototype system integrating data acquisition, behavior recognition, performance evaluation, and visualization modules was developed. Comparative experiments were conducted to assess system performance and evaluation capability. In terms of computational performance, the optimized model achieved a Top-1 Accuracy of 94.117% under image-based input conditions. The average inference time was 12.487 ms, and the model size was 71.298 MB, demonstrating strong real-time capability and deployment feasibility. Regarding behavioral evaluation, the mean score deviation was 0.66, the key action point recognition rate reached 98.62%, and overall system acceptance was 95.38%. Users rated the actionability of feedback suggestions at 4.88, indicating high practical value. Experimental results demonstrated that the proposed system outperformed existing methods in recognition accuracy, responsiveness, and user experience. The framework therefore provided an effective paradigm for the design and deployment of intelligent sports behavior evaluation systems in digital sports contexts. This study offered meaningful contributions to the fields of intelligent sports analytics, human behavior recognition, and interactive training systems.