GCANet: A Cross-Modal Pedestrian Detection Method Based on Gaussian Cross Attention Network

Pedestrian detection is a critical but challenging research field widely applicable in self-driving, surveillance and robotics. The performance of pedestrian detection is not ideal under the limitation of imaging conditions, especially at night or occlusion. To overcome these obstacles, we propose a cross-modal pedestrian detection network based on Gaussian Cross Attention (GCANet) improving the detection performance by a full use of multi-modal features. Through the bidirectional coupling of local features of different modals, the feature interaction and fusion between different modals are realized, and the salient features between multi-modal are effectively emphasized, thus improving the detection accuracy. Experimental results demonstrate GCANet achieves the highest accuracy with the state-of-the-art on KAIST multi-modal pedestrian dataset.