Confidence based active learning for vehicle classification in urban traffic

This paper presents a framework for confidence based active learning for vehicle classification in an urban traffic environment. Vehicles are automatically detected using an improved background subtraction algorithm using a Gaussian mixture model. A vehicle observation vector is constructed from measurement-based features and an intensity-based pyramid HOG. The output scores of a linear SVM classifier are accurately calibrated to probabilities using an interpolated dynamic bin width histogram. The confidence value of each sample is measured by its probabilities. Thus, only a small number of low confidence samples need to be identified and annotated according to their confidence. Compared to passive learning, the number of annotated samples needed for the training dataset can be reduced significantly, yielding a high accuracy classifier with low computational complexity and high efficiency. The detected vehicles are classified into four main categories: car, van, bus and motorcycle. Experimental results demonstrate the effectiveness and efficiency of our approach. The method is general enough so that it can be used in other classification problems and domains, e.g. pedestrian detection.