Abstract
In response to the challenge of limited accuracy in skeleton-based action recognition algorithms due to missing key points in complex environments under coal mines, enhancements were made to the Info-GCN++ model architecture. We proposed a miner action recognition model named ANODE-GCN, which integrated neural ordinary differential equations (NODE) with graph convolutional networks (GCN). The model predicted future motion sequences by analytically solving NODE in a dimensionally upgraded ODE space and combined these predictions with the actual observed motion states, thereby enhancing the recognition robustness of the model in handling partially missing skeleton sequences. Additionally, we designed a graph convolutional network SC-GC that integrated self-attention and coordinate attention mechanisms to differentiate between similar motion sequences in distinct actions. Ultimately, the miners' basic actions identified were correlated with environmental information to recognize more complex violation behaviors accurately. Experimental results demonstrated that on the public dataset NTU RGB+D120, with skeleton sequences completeness of 40% and 60%, accuracies of 71.96%/78.93% and 77.43%/81.29% were achieved, respectively, based on X-Sub/X-Set evaluation protocols. Ablation experiments based on the X-Sub evaluation protocol indicated that ANODE-GCN had an AUC of 67.13%, 10.75% higher than the Info-GCN++ baseline. On a self-built dataset, ANODE-GCN achieved an action recognition accuracy of up to 89.12% on the low-quality skeleton action test set. When the action information was matched with the environmental information, the average accuracy of miners' violation behavior recognition reached 91.7%, which was 6.7% higher than Info-GCN++.