As vehicles improved with technological advancements, faster speeds were attained and more serious accidents were caused. Vehicle accident statistics disclose that the main threats a driver is facing are from other vehicles. Consequently, developing on-board driver assistance systems enabling vehicle collision avoidance and mitigation has attracted more attention. In these systems, vehicle detection plays a key role. The most common approach to vehicle detection is using sensors such as radars and laser. However, they cost high to develop. Optical cameras, on the other hand, offer a more affordable and reliable solution. Visual information can be obtained without requiring any modifications to vehicles or road infrastructures. In this work, we will consider the problem of on-road vehicle detection from rear views of static images.
Several variables have made on-road vehicle detection challenging. First, vehicles typically come into view in different speeds and may vary in size, shape and color. Second, the appearance of a vehicle depends on its orientation and could possibly be affected by nearby objects which cast shadows on it. Third, the landscape along the road changes continuously while the illumination condition is unpredictable.
I propose to use Gabor filters for feature extraction and SVMs for detection. I believe that Gabor features are more appropriate in the context of our application since it provides a mechanism for achieving some degree of invariance to global illumination, selectivity in scale and orientation. Central to our approach is the idea of using Gabor filter banks to extract edge and line features at different scales and orientations. These features encode the coarse structure of a vehicle and can handle within-class variations. Thus, the statistics of these features could be efficiently used for SVM classifier training and vehicle verification.
The dataset will be collected from MIT CBCL car database, containing 516 rear and front view cars of size 128x128 pixels. Those images were caught during different time and different places so that a good variation of data can be ensured. The training set will contain a total of 500 images, both rear vehicle views and non-vehicles are included.