Adapting Deep Neural Networks for Pedestrian-Detection to Low-Light Conditions Without Re-Training

Vedant Shah, Anmol Agarwal, Tanmay Tulsidas Verlekar, Raghavendra Singh; Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops, 2021, pp. 2535-2541

Pedestrian detection is an integral component in many automated surveillance applications. Several state-of-the-art systems exist for pedestrian detection, however most of them are ineffective in low-light conditions. Systems specifically designed for low-light conditions require special equipment, such as depth sensing cameras. However, a lack of large publicly available depth datasets, prevents their use in training deep learning models. In this paper we propose a pre-processing pipeline, which enables any existing normal-light pedestrian detection system to operate in low-light conditions. It is based on a signal-processing and traditional computer-vision techniques, such as the use of signal strength of a depth sensing camera (amplitude images) and robust principal component analysis (RPCA). The information in an amplitude image is less noisy, and is of lower dimension than depth data, marking it computationally inexpensive to process. RPCA processes these amplitude images to generate foreground masks, which represent potential regions of interest. These masks can then be used to rectify the RGB images to increase the contrast between the foreground and background, even in low-light conditions. We show that these rectified RGB images can be used by normal-light deep learning models for pedestrian-detection, without any additional training. To test this hypothesis, we use the 'Oyla Low-Light Pedestrian Benchmark' (OLPB) dataset. Our results using two state-of-the art deep learning models (CrowdDet and CenterNet) show: a) The deep models perform poorly as pedestrian detectors in low-light conditions; b) Equipping the deep-networks with our pre-processing pipeline significantly improves the average precision for pedestrian-detection of the models without any re-training. Taken together, the results suggest that our approach could act as a useful pre-processor for deep learning models that aren't specially designed for pedestrian-detection in low-light conditions.