Study on the effect evaluation of physical education teaching based on situational perception and multi-scenario fusion recommendation algorithm

Abstract

Scenario perception enables personalized service recommendations by analyzing user contextual data. To enhance physical education evaluation, we integrate scenario-aware recommendation algorithms (e.g., TextRank with temporal weighting) to optimize real-time data filtering for teaching scenarios. Concurrently, we develop multi-object tracking models (e.g., twin networks with feature fusion) to address challenges in athlete movement analysis, creating a unified framework for data-driven teaching assessment. Second, the continuous word bag (model and user emotional tendency analysis algorithm on Baidu AI open platform), which can significantly reduce or eliminate the influence of the above data on the recommended content. Multi-objective tracking method based on the fusion of pedestrian derecognition and player number features. According to the problem of high target similarity, a feature dimension fusion strategy is proposed, which combines target characteristics from different dimensions (e.g., appearance, motion, and jersey numbers) to enhance the algorithm’s discriminative ability for similar targets. The experiments show that the proposed multi-feature fusion tracking method can still achieve a better tracking effect in the case of high target similarity, with direct applications in real-time analysis of athletes’ movement trajectories and tactical behaviors during physical education classes. Different IoU thresholds can ultimately affect the tracking performance of the algorithm to a large extent, providing quantitative support for optimizing teaching evaluation systems. It can be seen from the results that when the IoU threshold is 0.5, the algorithm achieves the best tracking performance, with the HOTA index improving by 3.8 points compared to the benchmark method, AssA index by 6.4 points compared to the benchmark method, and IDF 1 index by 7.1 points compared to the benchmark method. When the IoU threshold is 0.6 and 0.7, the HOTA index increased by 3.2 and 2.2 points compared to the benchmark method; the AssA index increased by 5.1 and 4.0 points with the benchmark method. To solve the problem of severe target occlusion, a twin network (a neural network architecture with two parallel branches for feature extraction and similarity calculation) is introduced to enhance anti-occlusion capabilities by modeling target motion and contextual information. The tracking method consists of three-level target associations, and the similarity needed to calculate the association is calculated using the information of twin network, multi-feature fusion, and spatial location. To solve the problem of severe target occlusion, the twin network method is introduced to use the algorithm in anti-occlusion. The experimental results show that the multilevel object tracking structure of the similarity calculation method of the twin network can effectively alleviate serious occlusion problems in sports events.