Performance Evaluation of YOLOv7 for Object Detection in Dark Environments

摘要

For a variety of applications, including security, surveillance, and robotics, the ability to detect objects in low light conditions is essential. Numerous algorithms have been modified to enhance performance in low-light conditions; however, the You Only Look Once (YOLO) algorithm is a standout example of a sophisticated computer vision technique, offering optimal runtime and precise object detection. In this work, we carried out an evaluation and comparison analysis on YOLOv7, focusing on low-light environment detection. To achieve this, we utilized the Exclusively Dark Image Dataset (ExDark) to train the YOLOv7 model, enabling it to identify twelve classes in dark environments. The results were presented and compared to previous research investigations that used YOLOv3, YOLOv5, and adaptive versions of YOLOv3 and YOLOv5 to detect in low-light environments. Notably, the precision and recall for YOLOv7 detection were measured at 0.7153 and 0.6556, respectively. YOLOv7 demonstrated a considerable improvement in mean average precision (mAP) when compared to prior research that used the same dataset.