A Large-Scale Object Detection Method for Complex Indoor Traffic Scenarios

In indoor environments, architectural traffic components such as stairs and steps are crucial for the safe passage of visually impaired individuals. However, due to spatial constraints, these structures appear as large-scale objects in object detection scenarios. Traditional object detection methods struggle to handle this issue due to limited effective receptive fields and insufficient feature representation capabilities. To address these challenges, this paper proposes a large-scale object detection model, SRASPP-YOLOv5, based on multi-scale feature fusion. Specifically, a serial residual multi-scale feature fusion module (SRASPP) is designed, which integrates standard convolution and dilated convolution in a serial structure to enhance the model’s perception of objects at different scales. Additionally, a Polarized Self-Attention (PSA) mechanism is introduced to improve feature ex-traction efficiency while maintaining computational complexity at a manageable level. Experimental results demonstrate that, compared to the baseline model YOLOv5 v6.0, SRASPPYOLOv5 improves accuracy metrics by more than 2% across different model sizes while maintaining an inference speed of no less than 50 frames per second. Notably, SRASPP-YOLOv5s achieves a 9% increase in the [email protected] metric compared to YOLOv5s, showcasing superior performance in complex indoor environments.