Improved Haar Wavelet Feature Extraction Approaches for Vehicle Detection

Feature extraction is a key point of pattern recognition. Wavelet features are attractive for vehicle detection because they form a compact representation, encode edges, capture information from multi-resolution, and can be computed efficiently. This paper focuses on the improvement of wavelet features. The wavelet features directly based on signed coefficients are easily affected by the varied surroundings and illumination conditions and cause high intra-class variability. In order to deal with this problem, three improved approaches based on unsigned coefficients are proposed. The results of these proposed approaches are compared with the current three methods. The proposed approaches show super performance under various illuminations and different roads (different day time, different scenes: highway, urban common road, urban narrow road).