Convergence and Divergence: A New Paradigm for Pedestrian Detection

Complex backgrounds, scale and occlusion variance have long limited the accuracy of pedestrian detection. In this paper, we propose a pedestrian detector named Convergence and Divergence (CADNet). In “Convergence” network, we propose a cross-scale semantic alignment block (CSAB). CSAB effectively mitigates the background interference and resolves scale variance through multi-scale global contexts aggregation, without extensive computational overhead. In “Divergence” network, we propose a receptive field differentiation block (RFDB) to tackle the challenges of scale and occlusion variance. RFDB generates discriminative features with varying receptive fields, effectively capturing pedestrians across different scales and occlusion conditions. Due to the effectiveness of the proposed components, CADNet achieves an excellent performance of 8.47% and 2.16% MR−2 on a Reasonable subset of CityPersons and Caltech, respectively. Extensive experiments demonstrate the robustness and efficiency of CADNet, ensuring its superior performance in various scenarios.