Perceptual Feature Integration for Sports Dancing Action Scenery Detection and Optimization

Deciphering the complex semantics within varied dancing sceneries is crucial for a multitude of AI endeavors. It can facilitate applications like dancing action optimization and dancing education. In our research, we propose a sophisticated approach to discerning multi-faceted perceptual visual features for accurately identifying dancing scenic imagery with intricate spatial designs. Our work centers on crafting a deep hierarchical structure adept at simulating human gaze patterns, utilizing the BING metric to pinpoint objects and their components within scenes at different scales. To emulate human visual dynamics, we introduce a Robust Deep Active Learning (RDAL) methodology, systematically creating gaze shift paths (GSPs) and capturing their profound representations. A key novelty of RDAL is its resilience to inaccuracies in labeling, employing a strategically designed sparse penalty framework that facilitates the exclusion of non-informative or irrelevant deep GSP attributes. Furthermore, we propose a manifold-regularized feature selector (MRFS) to isolate premium deep GSP features, concurrently developing a linear SVM for dancing scene recognition. Our method's efficacy, validated through rigorous testing, not only showcased its enhanced performance across conventional scenic datasets but also highlighted the exceptional discriminating power of deep GSP features in a specialized dataset for recognizing different dancing actions. Finally, the dancing actions can be optimized using a probabilistic model.