On-road vehicle detection with monocular camera for embedded realization: Robust algorithms and evaluations

Vehicle detection is critical operation in automotive active safety systems. Although there are a number of vehicle detection techniques available in literature, computationally efficiency for realization on embedded platforms is not explored and addressed in most existing works. In this paper, we present a computationally efficient vehicle detection algorithm that is particularly designed for architectural translation into efficient embedded hardware. The proposed method uses camera calibration to derive the appropriate window scales that must be used for vehicle detection, resulting in a computational cost reduction of over 10 times. In addition to reduction in sampling windows, the proposed vehicle detection technique uses a novel multi-part based vehicle detection method which detects the vehicles that pose the highest risk to the ego-vehicle. The proposed method is evaluated using different datasets and computational savings are seen in orders of magnitude as compared to conventional sliding window approaches, without compromising on accuracy.