Vehicle model recognition using geometry and appearance of car emblems from rear view images

In this paper a novel vehicle model recognition approach is presented modelling the geometry and appearance of car emblems (model, trim level, etc.) from rear view images. The proposed system is assisted by LPR and VMR modules. Thus, a generic methodology is defined to build a hierarchical structure of car-make-dependent vehicle model classifiers. The emblems location, size and variations are firstly learnt. Then, the appearance of each badge is modelled using a linear SVM binary classifier with HOG features and the outputs of each individual classifier are converted to an estimate of posterior probabilities. A specific probability is computed for each hypothesis (model) integrating the posterior probabilities of all the emblems using the geometric mean. Inference about the most probable car model is finally carried out selecting the model with the maximum probability. We evaluate this approach on a dataset composed of 1.342 images (910/432 for training/test) corresponding to 8 different car makes and 28 different car models (52 considering generations) achieving an overall accuracy of 93.75%.