On road vehicle detection by learning hard samples and filtering false alarms from shadow features

This paper proposes a two-step approach to detect preceding vehicles on the highway. The first step of the approach is to approximate vehicles' potential locations through search for shadow area of lower part of vehicle. To find these shadows, the Haar-like feature with AdaBoost was used to train a shadow detector offline. A re-learning process with hard training samples was applied to increase the detection rate. Additive Markov chain method was also applied to eliminate most false alarms. Shadow area detection was found to be fast enough to focus on interesting local regions that have high probability to include vehicles. In the second step, on the basis of previously detected shadow areas, we were able to set extended areas in which to perform vehicle verification. Proving the existence of vehicles is based on distribution of edge histograms in ROI and training by the SVM algorithm. The experimental results proved that the proposed system could be used robustly and accurately for real-time preceding vehicle detection.