Machine Learning for Identifying Pedestrians and Vehicles in Low-Light Images: A Survey

Pedestrian detection has applications in many areas, like autonomous vehicles, surveillance systems, and robotics. This is quite a challenging problem in real-world scenarios, for reasons such as illumination changes and diversity. Of the challenges presented in the literature, low illumination is clearly of great magnitude, as it can be affected by several factors related to time, geography, and weather. While Light Detection and Ranging is one of the classical sensor modalities used for distance measurement and is very accurate, it is not suited for gaining general environmental information. Among all state-of- the-art pedestrian detection approaches, which are numerous, hardly any consider lousy visibility. In low-light scenes, autonomous vehicles are seriously limited concerning the detection of object motion on roads, which increases the danger of accidents. Many technologies have been developed against these difficulties; among them, one of the most efficient ones is machine learning. They can learn complex features directly from data, which improves the accuracy of object detection under low illumination conditions. Other techniques, such as image enhancement, noise reduction, and transfer learning, have been infused into a deep learning pipeline to increase the system's robustness. The present paper surveys in detail state- of-the-art machine-learning models and methods for pedestrian and vehicle detection from low-light images. Synthesis of the recent trends demonstrates that machine learning has enormous potential and can significantly improve the accuracy and reliability of object detection systems working under various illumination conditions.