Fine-Grained Human Action Recognition via Laban Movement Analysis and Attentive Bi-LSTM

The study uses 3D bone sequences to investigate fine-grained classification of dance types. Two challenges exist, including modeling complex time dynamics and extracting semantic motion features. Unlike those methods that rely on low-level joint coordinates or only capture global time dependence, we propose to encode based on the high-order motion semantics of Laban motion analysis (LMA) and dynamically focus on differentiated time fragments. First, the original bones will be converted into a standardized coordinate system to improve robustness. Then, we extract LMA descriptors, which cover body posture, force (dynamics), and spatial relationships to form a rich representation of the characteristics of each frame. We introduced a two-way short- and long-term memory network (Attention-BiLSTM) with enhanced attention. The two-way short- and long-term memory network encodes the LMA feature sequence in two time directions, while a self-attention module assigns weights to the frame, emphasizing key motion fragments that are critical for type recognition. Experiments on the challenging AIST++ dataset verified the effectiveness of each component. Comparative analysis with the LSTM-based benchmark model and state-of-the-art methods shows that the model performs better. The visualization results, containing category activation maps, feature correlation maps, and attention heat maps, offer easy-to-understand modes. Action patterns with rich semantics and differentiation are used for each dance type in the model.