Spatial-temporal attention augmented graph convolution method based on human posture response

Abstract

The requirements for the timely response of on-site lighting in stage performances or smart classrooms are becoming increasingly refined. Existing studies have neither focused on real-time human pose estimation in stage performances nor lacked the integrated integration of real-time human motion analysis and lighting control mechanisms. To this end, we propose a spatial-temporal graph convolutional neural network (ST-E-GCN) that integrates an enhanced efficient channel attention mechanism (EECA). By establishing a dynamic mapping function between joint angular velocity and channel weights, we solve the bottleneck of insufficient decoupling of spatial-temporal features in traditional skeleton action recognition models in dance movements such as high-speed rotation. This model innovatively introduces a motion dynamics perception mechanism, increasing the feature weights of centrifugal joints to 2.3 times the benchmark value and significantly enhancing the discriminative representation of key motion joints. Based on this, an embedded control system was designed. Through the optimized DMX512 protocol parsing, the ST-E-GCN classification results were mapped to multi-channel PWM signals to achieve lighting control. The experimental results show that on the AIST dance dataset, our method improves the movement classification accuracy by 4.87 % compared to the baseline. The hardware system can achieve precise synchronization of lights and performers' actions. It meets the real-time response and precise adjustment requirements for lighting control.