Vehicle Detection in Monocular Night-Time Grey-Level Videos

Abstract

Road traffic accidents have been a huge problem for many decades. Ongoing efforts aim at the development of systems that can help to minimize the number of fatal accidents by providing warnings to the driver. These systems are known as Vision-Based Driver Assistance Systems. There have been many inventions since the 1980’s to make driver assistance systems more reliable. Many researchers contributed their efforts to making such systems more accurate. Driving under any conditions involves many risks on busy roads. Countless means have been proposed to reduce the risks and prevent accidents. One such approach is motion detection of nearby moving vehicles using different image analysis algorithms. Vision-based driver assistance systems are an amalgamation of technology, software programs, hardware, and GPS. The number of vehicles on roads increases every year. Unfortunately, the numbers of accidents are also proportional to this increase. There have been many safety measures undertaken by the joint collaboration between researchers and the automotive industry to include vision-based systems and increase road safety. Vehicle detection at night-time is more difficult than detection at day-time due to the limited availability of features and more challenging lighting conditions. When driving at night-time, vehicles approaching from the front are only visible by their headlights. This thesis presents a night-time vehicle detection system based on the processing of gray-level video. An automatic threshold step is introduced, which reduces noise caused by non-candidate objects (such as street lights, traffic lights, road sign boards and so forth). Regions of interest (RoI) are identified by training a classifier for this task. In the described experiments, RoIs are headlights of approaching vehicles from the front. A large dataset has been used to train the classifier. The thesis proposes a specific technique to ensure time-efficiency when training the classifier. Experiments evaluate a monocular vision system capable of detecting vehicles approaching from the front. The proposed system uses a Haar-like feature approach for night-time gray-level video sequences. The approach detects vehicles at nighttime by searching for headlights (i.e. the RoI) in recorded frames. Experiments demonstrate the effectiveness of the proposed system. Keywords: Vision-based driver assistance system, vehicle detection, monocular vision, Haar-like features, headlights, night-time vision.