Real-Time Forward Collision Warning System Using Nested Kalman Filter for Monocular Camera

With the rise of autonomous cars and advanced driver assistance systems, the demand for Forward Collision Warning to support the Advanced Driver Assistance System (ADAS) is increasing and most of these systems use many sensors such as Iidar, Global Positioning System (GPS), radar, stereo cameras which are big and expensive. Such sensors are not readily available for Vulnerable Road Users (VRU) such as cyclist, motorcyclist and Personal Mobility Devices (PMDs) users as they mostly only have a monocular camera on their smartphones or dash camera for navigation and recording for safety. Hence, this paper aims to create a real-time Forward Collision Warning system (FCWs) for VRU using only a monocular camera. However, this is a challenge as the prediction of distance, velocity and time to collision for the FCW is noisy and not accurate without sensor fusion involving other sensors. Moreover, existing vehicle detection algorithm such as Support Vector Machines (SVM) have challenges in real-time application. To address these limitations, this paper proposes a method to obtain a real-time stabilize Forward Collision Warning system to avoid false alarm based on a monocular camera without additional sensors. This is achieved using a nested Kalman filter to first predict and stabilize distance detected and subsequently, predict and stabilize velocity of vehicles detected as well as time-to-collision on a monocular camera platform such as handphones and other consumer devices based on You Only Look Once (YOLO) vehicle detection algorithm.