A Multi-Branch Anchor-Free Detection Algorithm for Hospital Pedestrian

Accurate and efficient real-time pedestrian detection in hospitals is crucial for improving safety operations and effective management. However, this task poses significant challenges due to complex scenes, dense crowds, and pedestrian occlusions. This paper proposes a multi-branch anchor-free algorithm for hospital pedestrian detection. Firstly, a multi-branch network structure is added after the backbone network of the model to adapt to multiple key local features of pedestrian targets. Subsequently, a distance loss function between key regions is designed to guide the branch networks in learning the differential detection positions of pedestrians locally. Furthermore, using ResNet34 as the baseline feature generation network, four upsampling blocks are appended at the end to form a hourglass structure, enhancing the branch network's understanding of spatial information in pedestrian local features. Lastly, a local feature selection network is proposed to adaptively suppress non-optimal values from the multi-branch outputs, eliminating redundant feature boxes during prediction. Experimental results demonstrate that the method achieved an AP of 89.2% on the CrowdHuman dataset, indicating high detection accuracy. Additionally, on the HospitalPerson dataset for pedestrian detection in hospitals, the F1 score, Recall, and AP reached 0.9, 94.59%, and 85.42% respectively, showcasing the superior performance of the proposed method in hospital pedestrian detection, particularly in crowded and heavily occluded scenarios.