Enhancing Visibility for Autonomous Vehicle Object Detection in Dust or Sandstorm Conditions Using Dark Channel Prior and YOLOv9
Mir Mohammad Ali, Mahmudul Hasan, Meherun Nesa Meem, Feroza Naznin, Riad Hassan, Md Zahidul Islam · 2024
The detection of objects by autonomous vehicles is hampered by unfavorable weather conditions such as dust or sandstorms. These conditions significantly reduce visibility up to reduce visibility significant and blur object features, creating notable challenges for traditional object detection models. Existing models often fail to perform effectively under such conditions due to decreased image clarity and increased noise, presenting a critical gap in current research. To overcome this challenge, we proposed a new image enhancement method based on Dark Channel Prior (DCP) to enhance images in dust or sandstorm weather. For image enhancement, our methodology employs a multi-step process: we first estimate airlight and initial transmission using haze map transmission, followed by morphological operations and adaptive weight adjustments via Kirsch filters. Our algorithm introduces several parameters that are absent in DCP, including regularize-lambda, sigma, and delta, providing with greater control over dehazing quality. These parameters optimize the balance between preserving structural details and achieving effective haze removal across diverse scenes. Then the haze-free image is fed to the state-of-the-art You Only Look Once (YOLOv9) object detection model to detect objects such as persons,cars, bikes, buses and trucks. Our methodology was evaluated using the DAWN dataset, where the enhanced images achieved a Peak Signal-to-Noise Ratio (PSNR) of 16.44 and an Average Gradient (AG) of 321.93, outperforming other methods. The proposed method represents a notable advancement in enhancing image clarity and detection in adverse weather condition.