Effectiveness evaluation of sprint sports techniques and tactics based on deep learning

The assessment of sprint-like sports techniques and tactics involves evaluating their underlying running principles, structures, functions, methods, and real-world applications using existing technology. Particularly in the context of the fast-evolving sprint domain, a deficiency in evaluation expertise can result in lagging behind. Using deep learning algorithms, useful features, and information can be automatically extracted from extensive training and competition data, providing a more accurate and objective basis for evaluating sprint sports techniques and tactics. We can reduce manual intervention and tedious workloads by applying deep learning algorithms in automated processing and analysis, enabling intelligent training and competition. Deep learning algorithms can uncover and analyze patterns and trends in data, assisting coaches and athletes in identifying potential problems and areas for improvement and providing a foundation for athletes to develop targeted training plans and competition strategies. Therefore, studying the role of Deep Learning (DL) in the technological process is of great significance. This paper aims to effectively combine DL algorithms with sprint skill evaluation, utilizing DL algorithms for mining and leveraging them to achieve optimal sprint skill performance through effective tactical evaluation. This study trains a new model based on the classic VGGNet-16 model and compares different CNN models. The experiment reveals that the CaffeNet model is more susceptible to objects that do not entirely contain the target, resulting in an enlarging search box and eventually losing track of the target. The experimental results demonstrate that maintaining a step frequency of approximately 95% is crucial for achieving the highest speed. The SDP algorithm was tested on the VOC2007 dataset and achieved good results, with a mean average precision value of 69.4%, significantly surpassing the benchmark Fast R-CNN algorithm. Additionally, we have implemented a traditional tracking algorithm, the CamShift algorithm, based on color features.