CNN Combined With a Prior Knowledge-based Candidate Search and Diffusion Method for Nighttime Vehicle Detection

Driver assistance systems or smart headlamp technology require accurate vehicle detection in nighttime traffic environments. This study proposes a novel nighttime vehicle detection (NVD) algorithm that can be applied to these technologies, and the algorithm was implemented using image processing. Images taken on roads at night are basically dark and lack information regarding vehicle appearance. Additionally, they contain significant noise caused by various lights and the scattering or reflection of these lights. Therefore, it is difficult to increase the performance of existing NVD methods. In addition, recent end-to-end convolutional neural network (CNN)-based object detection (OD) methods exhibit low NVD performance owing to their poor learning capability caused by lack of information and noise in the images. This study presents new methods to overcome the limitations of NVD implementation: 1) We propose a candidate search and diffusion method based on the use of experimental heuristics and hand-crafted features to utilize the characteristics related to light emitted from vehicle headlamps or taillamps. 2) We propose a CNN method that combines the approaches applied in latest CNN-based OD method with the proposed candidate search and diffusion method. To demonstrate the superiority of the proposed method, we conducted experiments and compared the proposed method to recent CNN-based OD methods. The experimental results demonstrated the higher detection performance of the proposed method compared to other methods. The code is available at https://github.com/sjg918/gj-nvd-diffusion/