You Only Look Once in Dark: An Analytical Approach for Low Light Object Detection

In recent decades, object detection has been one of the most significant areas of research, which enables an autonomous machine to make intelligent decisions without any human intervention. Low-light object detection faces numerous challenges including noise, granules, and ambiguity in no or limited color information. Single Shot Detectors (SSD), especially, You-Only-Look-Once (YOLO) one of the latest improvised object detection techniques, find objects within the image in a single forward pass instead of separating into two stages like Faster R-CNN by identifying regions of interest. This paper conducts an extensive analysis of the performances of the latest ML-mediated object detection technique, YOLO. All the versions of YOLO are evaluated based on various ranges of lux value in an experimented environment. The main focus of the research work is to compute the performances of YOLO techniques in varying light conditions starting from bright to low-light and darkest environments. The analysis presents a compatible result within a considerable threshold between low and high lux values $(\text{lumen}/m^{2})$. The real-world experiment is established on a standard dataset, COCO, under varying light-conditioned setups where the vision system frequently transits from brighter to darker or vice versa. The experimental computation reveals the chronological performance order: YOLO v8$210-600 \ \ \text{lm/m}^2$ and $30-210\ \ \ 1\mathrm{m}/ {m}^{2}$ lux values in consecutive brighter and darker environments. A dynamic Decision Support System(DSS) has been introduced that confirms a balanced performance of object detection in varying light conditions.