Robust vehicle detection is a challenging task given vehicles with different types, and sizes, and at different distances. This paper proposes a Boosted Gabor Features (BGF) approach for vehicle detection. The two main conventional Gabor filter design approaches are a filter bank design approach with fixed parameters even for different applications and a learning approach. In contrast, the parameters of our boosted Gabor filters, learned from examples, differ from application to application. Moreover, our boosted approach optimizes the filter parameters for every image sub-window, and the boosted filters have a large response for sub-windows containing a part of a vehicle resulting in a greatly improved performance in vehicle detection.

Our vehicle detection has two basic phases in which we build a multi-resolution hypothesis-validation structure. In the vehicle hypothesis generation phase, hypothesis lists are generated for three ROIs with different resolutions using horizontal and vertical edges ,and following that, a hypothesis list for the whole image is obtained by combining these three lists. In the subsequent hypothesis validation phase, we validate the vehicle hypothesis list by inputting the boosted Gabor feature vector into the support vector machine.

In the context of vehicle detection, the resulting system yields detection rates comparable to the best previous systems while achieving a 20 frames per second real-time performance on a Pentium(R)4 CPU 2.4GHz.