Human risky behaviour recognition during ladder climbing based on multi-modal feature fusion and adaptive graph convolutional network

Human falls during ladder climbing are typically instantaneous, making the timely and accurate determination of security risks during ladder climbing a challenging engineering issue. A skeleton-based behaviour recognition method is proposed for the real-time detection of human risky behaviour during ladder climbing in construction scenes. This method used a multi-modal feature fusion strategy to enhance the semantic information of the skeletal data and used an improved adaptive graph convolutional network by partial dense connection for behaviour recognition. This method was evaluated through ablation and comparative experiments on the public behaviour dataset. The results demonstrated its advantages in balancing accuracy and model complexity. Moreover, the experiment results on the ladder climbing behaviour dataset also validated its effectiveness in practical applications. The method in this paper will hopefully help to guarantee the personal safety of construction workers and provide information-based advance warning of security risks during ladder climbing.