In sports teaching, the reliance on manual visual assessment of students' movement quality leads to subjectivity and low efficiency. This paper proposes an automatic assessment algorithm for sports movement quality based on deep learning. Firstly, a multimodal data platform is constructed to obtain RGB videos, depth skeletal sequences, and inertial measurement unit data. After time alignment and normalization, a three-dimensional skeletal motion sequence is generated. An adaptive spatio-temporal graph convolutional backbone network is designed. It dynamically models the collaborative relationships between non-adjacent joints via a learnable adjacency matrix and introduces a temporal multi-scale aggregation module to capture movement patterns at different time granularities. Furthermore, a hierarchical scoring regression module is proposed, which includes a global-local joint encoder, a segmented fine-grained scoring head, and a bias correction branch based on contrastive learning. It jointly models overall coordination, local posture, and individual difference compensation. Experiments on a self-built sports movement quality dataset show that the algorithm achieves a mean absolute error of 3.62 points, a Spearman rank correlation coefficient of 0.873, and a parameter count of 3.72 million. After pruning and TensorRT optimization, it performs real-time inference at 28.6 ms per sample on a Jetson Nano device. Ablation experiments and robustness tests verify the effectiveness of each core module, and the proposed method outperforms existing mainstream models in terms of accuracy and efficiency, providing a feasible technical solution for automatic movement assessment in sports teaching scenarios.