CLS‐Transformer for Home Service Robots: A Lightweight Action Recognition Model Based on Skeleton and Native Length Encoding

Home service robots have the capability to monitor and perform statistical analyses of daily activities through action recognition techniques. For autonomous mobile home service robots with limited computing power, it is crucial to develop lightweight and highly accurate action recognition networks. The paper presents a native length encoding approach alongside a human skeleton‐based action recognition model (named CLS‐Transformer). The model incorporates a novel cue coding strategy that captures richer feature information by accounting for variations among different skeletal joints within the same frame as well as changes in identical joints across consecutive frames. To minimise redundant dense skeleton computations and eliminate processing of irrelevant padded video frames, the CLS‐Transformer's spatiotemporal encoder integrates both spatial and linear multi‐head attention mechanisms. Due to poor generalisation and robustness of existing models, a contextual attention mechanism is employed to enhance feature aggregation. CLS‐Transformer is evaluated on three datasets: NTU RGB+D 60, NTU RGB+D 120 and HS‐AR (a self‐constructed home service action recognition dataset). Related ablations are also done to demonstrate the effectiveness of each key component in CLS‐Transformer. The experimental results demonstrate that CLS‐Transformer achieves superior recognition accuracy and inference speed compared to most current methods, while maintaining lower computational complexity.