Cross-task feature alignment for seeing pedestrians in the dark

Abstract Pedestrian detection in low-light environments has been an extremely challenging task because of the serious degradation of color and texture information. In the latest research, multi-task learning is introduced into image relighting and pedestrian detection tasks, which improves the performance of detecting pedestrians in low-light environments significantly. However, many problems in the multi-task learning period, including the misalignment of scale and channel of features from image relighting and pedestrian detection tasks, remain unresolved, thereby resulting in insufficient feature representation. In this paper, we propose a novel cross-task feature alignment method to tackle the aforementioned problems. Specifically, the proposed method imposes four feature alignment (FA) layers before the feature fusing and sharing step in multi-task learning period to align the scale and channel of features across tasks and fine-tune feature representation iteratively. In addition, we design a novel multi-scale feature-enhanced detection network to further improve the performance of the detector. Experimental results from simulated and real-world scenarios prove that our method prominently boosts the ability to detect pedestrians in the dark.