Object classification in traffic scenes using multiple spatio-temporal features

Object classification is a widely researched area in the field of computer vision. Lately there has been a lot of attention to appearance based models for representing objects. The most important feature of classifying objects such as pedestrians, vehicles, etc. in traffic scenes is that we have motion information available to us. The motion information presents itself in the form of temporal cues such as velocity and also as spatio-temporal cues such as optical flow, DHOG [6], etc. We propose a novel spatio-temporal feature based on covariance descriptors known as DCOV which captures complementary information to the DHOG feature. We present a combined classifier based on properties of tracked objects along with the DHOG and the DCOV features. We show based on experiments on the PETS 2001 dataset and two video sequences of bicycle and pedestrian traffic that the proposed classifier provides competent performance for distinguishing pedestrians, vehicles and bicyclists. Our method is also adaptive and benefits from the availability of more data for training. The classifier is also developed with real-time feasibility in mind.