Vehicle type recognition with match refinement

We describe a system for automatic recognition of vehicle type (make and model) from frontal views, aimed at secure access, surveillance and traffic monitoring applications. The system extracts gradient features from reference patches in images of car fronts and performs recognition in two stages. In the first stage, gradient based feature vectors are used to produce a ranked list of possible candidate classes. The result is then refined by using a novel match refinement algorithm that maximises the discrimination between the subset of most likely classes by optimising for object pose and adaptively normalising feature vectors. We test the system on over 1000 images containing 77 difference vehicle classes, and demonstrate that such a system can provide reliable verification (EER<3.8%) and identification (P_id=94.4%) of vehicle type.