Urban road user classification framework using local feature descriptors and HMM

Surveillance and safety systems for pedestrians and bicyclists are becoming much more important because there continue to be a large number of traffic accidents that involve vulnerable road users. In this paper, we propose an urban road user classification framework using local feature descriptors and hidden Markov models (HMM). Our framework achieved pedestrians, bicyclists, motorcyclist classification in high accuracy. The framework consists of two classification methods: pedestrian-bicyclist classification and bicyclist-motorcyclist classification. First, we discriminate between pedestrians and bicyclist-like objects using histograms of oriented gradients (HOG)-based classifiers. We implemented a cascade classifier using generic HOG and our original local feature descriptor called co-occurrence semantic HOG. Bicyclist-like objects mainly consist of bicyclists and motorcyclists. We focused on the objects´ leg motions and classify them using the hidden Markov models (HMM)-based motion models. We conducted experiments with real traffic scenes to evaluate the performance of our framework. The experiments for pedestrian-bicyclist classification and bicyclist-motorcyclist classification are conducted independently and both methods achieve nearly 90% on classification.