Detecting Multi-depth Headlight of Traffic Videos in Nighttime Based on Costructing Model according to Atmospheric Scattering Principle

According to the difficulties to extend the detection distance of the traffic video at night and simultaneously to increase the headlights detection rate, we propose a novel headlight recovery model based on the atmospheric scattering principle. We firstly analyze all light sources in nighttime traffic scene, then build a headlights recovery model based on the atmospheric scattering principle. After redefining the ambient light, we present new methods to estimate the unknown parameters, such as transitivity, ambient light, scene depth, etc., the recovery results only including headlights and small numbers of reflection lights on the road surface. Finally, the video scene is partitioned to three parts: far, middle and near regions. The recovery results are filtered according to the headlight geometrical characteristic to increase the headlight detection rate. The results on 9 videos, totally 14 492 frames, show that we have extended the detection distance, and at the same time, the average detection rate is increased by 31.39%, the average missing and false detection rates are decreased by 20.93% and 10.46% respectively, compared with other advanced algorithms.