Vision based preceding vehicle detection using self shadows and structural edge features

An innovative approach for on-road real-time preceding vehicle detection system is presented in this paper. Vehicle detection is performed by using knowledge based candidate generation followed by appearance based verification. The primary shadow present underneath the vehicle chassis i.e. self shadow is used to generate candidate regions in the image. The use of only self shadow provides improved results and robustness as compared to cast shadows utilized in other approaches. The vehicle class has large intra-class variance due to which a large training dataset with normalized samples is needed for accurate classifier design. It is proposed that the deterministic structure of the contour of vehicles remains same irrespective of its appearance. Hence, structural analysis using the edge based features can be used for classification. It is proposed that a smaller training data-set which is not necessarily normalized is sufficient for good classification results using this analysis. This leads to reduced complexity in system design.