Semi-automatic training of a vehicle make and model recognition system

We propose a system for vehicle Make and Model Recognition (MMR) that automatically detects and classifies the make and model from a live camera mounted above the highway. Our system consists of a vehicle detection and MMR classification component. The vehicle detector is based on HOG features and can locate 98% 98% of the vehicles with minimum false detections. We use a Convolutional Neural Network (CNN) for MMR classification on the vehicle locations. We propose a semi-automatic data-selection approach for the vehicle detector and the MMR classifier, by using an Automatic Number Plate Recognition engine for annotating new images, requiring minimal human annotation effort. In our results we show that our MMR classification has a top-1 accuracy of 98% 98% for 500 vehicle models, where more than 500 training samples per model are desired to obtain accurate classification