An enhanced relation-aware global-local attention network for escaping human detection in indoor smoke scenarios

Accurate and rapid human detection is crucial for emergency rescue in indoor fire scenarios, and surveillance video provides more possibilities for this work. However, smoke from fires reduces the visibility of humans in surveillance videos, which weakens the expression of human characteristics and makes it difficult to detect humans. To the best of our knowledge, existing studies do not include detecting human in smoke. However, human rescue often occurs in smoke scenarios in real-world settings. We introduce a new problem for human detection in this paper and propose an experimental investigation based on an enhanced relation-aware global–local attention network for escaping human detection in indoor smoke scenarios. First, we address the relationship of humans in smoke scenarios between the global body and local part based on a relation-aware global–local attention module, which can solve the problem of human disturbance by flowing smoke. Second, a prediction architecture considering interference is built to address human escape postures and video scales to obtain better human detection results. In addition, our method considers a faster frame rate for effective deployment. Finally, we propose a data simulation strategy for escaping humans in indoor smoke scenarios and establish a human detection dataset to prove the validity of our method. Accuracy and speed are used as evaluation criteria for the two testing datasets. Experimental results show that it is feasible and effective for human detection in smoke scenarios.