Deep learning is an important path for the digital reform of college sports teaching. In this context, the study is first conducted to acquire sports data in colleges and universities. Openpose technology is used to collect skeletal point coordinate data and process it to construct a human sports model. Then, based on the similarity comparison of sports poses, the generated sports sequences are filled to simulate the real human sports sequences and data, and the DTW algorithm is used to match the sports sequences and data with the features, eliminate the inconsistency between the length of the poses to be matched and the template pose sequences, and then give the scores of the poses to realize the evaluation of the learners’ sports performance. Combining the principles and functions mentioned above, we construct a system for evaluating sports teaching and analyze its application examples. The distribution of the distance between the quaternion-based swing action and the standard swing action is within the range of 0.18~0.25, and the distance between various types of actions and the standard action is not large. The professional scoring was 5.199 points higher than the DTW-based stance score of 90.183 points. In addition, in the experimental group and the control group, there is a significant difference in the analysis of students’ movement skills assessment, P<0.05, indicating that compared with the traditional teaching mode, the deep learning-based physical education teaching system has a strong relevance as well as the ability to provide real-time movement feedback and movement correction for students’ movement practice.