Global context-aware and attention mechanism method for small-scale pedestrian detection

Small-scale pedestrian detection is a major challenge due to limited pixel resolution and insufficient distinguishing features, frequently resulting in incorrect or missed detections. To address it, this paper proposed a global context-aware and attention mechanism algorithm for small-scale pedestrian detection. Firstly, considering the problem of small-scale pedestrian features gradually decreasing with network depth, we leverage the advantage of Transformers in capturing long-range dependencies. This allows us to design a global context information module that can retain a large number of small-scale pedestrian features. Then, considering the issue of small-scale pedestrian features easily being confused with background information, a Coordinate and Channel Attention Module (CCAM) is proposed. Coordinate attention can capture direction-aware and position-sensitive information, which helps the model to locate and recognize objects of interest more accurately. Channel Attention can effectively enhance small-scale pedestrian features and suppressing background information. Experimental results on the CrowdHuman dataset fully demonstrate that the proposed method can significantly improve the detection ability for small-scale pedestrian.