Vision-based Nighttime Vehicle Recognition for Smart Headlamps

Recently, there has been an increasing need for algorithms capable of precisely and rapidly recognizing vehicles at night via smart control of headlamps. In this study, we constructed an algorithm that could detect vehicles approaching the main vehicle and vehicles moving in the same direction as the main vehicle through images taken in front of the vehicle on a road at night. The algorithm mainly involved 1) the generation of vehicle candidates (VCs), 2) the classification of VCs, and 3) the tracking of VCs. VC generation generally begins with the extraction of light-blobs from an image and the pairing of these blobs. However, because various lights are mixed, it is difficult to identify which of these lights originate from vehicles. To solve this problem, we constructed multiple feature maps that are likely to closely relate to the light emitted from head and tail lamps and calculated the stereo disparity. The feature maps and stereo disparity were used for light-blob pairing to generate VCs. Subsequently, VC classification and tracking were performed. VC classification was performed using a convolutional neural network. The classifier indicated with probability whether the VC was a vehicle approaching the main vehicle, a vehicle going in the same direction as the main vehicle, or a non-vehicle. VC tracking performed via a Kanade−Lucas−Tomasi-based feature tracker enabled robust vehicle detection between consecutive input images. We showed that the proposed algorithm can be applied to the control of smart headlamps through real vehicle experiments. © ICROS 2024.