Pedestrian detection based on vision-language semantics with global adaptive adjustment

Pedestrian detection is the primary task of automated driving and intelligent video surveillance systems. Context-Aware Pedestrian Detection via Vision-Language Semantic Self-Supervision (VLPD) greatly improves the detection accuracy of single-stage pedestrian detectors. Meanwhile, to maintain reasoning speed, VLPD adopts ResNet-50 as its backbone network, which undoubtedly poses a significant limitation for single-stage detectors that require direct category prediction and bounding box regression on feature maps. To tap into the potential of CNNs in representation capability, we propose a novel simplified architectural unit, the Channel and Spatial Global Pooling Attention Module (GPA), which integrates activation channels and spatial weights attention maps through parallel computation to achieve adaptive feature refinement of backbone output feature maps. Furthermore, we optimize the module structure of the VLPD self-supervised prototype semantic contrast method, significantly enhancing the detector’s ability to discriminate and detect pedestrians in complex urban street environments. With only a 0.2FPS decrease in reasoning speed, the miss rates on the Heavy Occlusion subsets and Reasonable subsets of the Citypersons dataset are reduced by 2.41% and 0.72%, respectively, achieving state-of-the-art (SOTA) performance for single-stage detectors on this dataset. On the Heavy Occlusion subset and the All subset of the Caltech dataset, the performance decreased by 2.90% and 0.80%, respectively. Without using additional data, this method can rival the detection accuracy of two-stage detectors.