Scene-Adaptive Real-Time Fast Dehazing and Detection in Driving Environment

Real-time and effective dehazing is crucial to ensure safe and smooth operations of driving in foggy conditions. In this paper, a novel vision-based scene-adaptive real-time dehazing method for continuous video frames is developed, and the enhanced frames are fed into a detector in order to assess the value of image enhancement for detection tasks. The defogging stage is established by importing scene adaptation in order to improve the effectiveness of continuous frames defogging. Considering the distribution of atmospheric light intensity within the frame, the positions of the frame's key points are determined in order to distinguish the scene differences between frames. Then, an algorithm is proposed for the inter-frame perception of atmospheric light intensity based on key points ("inverted triangle"), which employs a dynamic updating to adjust the estimated update frequency of atmospheric light intensity in time based on the scene change. Further, frames are inverted to estimate the transmission map based on the scattering model, which can quickly recover free of fog. Finally, the adaptive region of interest is set based on the obtained key point of frames, and the enhanced image is fed to the detector. The experimental results indicate that the proposed method can effectively handle sudden scene changes, as the defogging accuracy can reach 98.58%, the average defogging time of each frame is only a few milliseconds, and the detection accuracy can be improved by 4.11%.