Low-resolution human detection and gait recognition in natural scenes

Fast and stable detection of humans in natural scenes is a challenging task due to the varying appearance of the target and diverse background. Under these circumstances a higher-level analysis, e.g. action classification, becomes even more difficult, as it is dependent on the quality of the preceding steps in the processing pipeline. In this paper we address this issue on both levels: low-level detection, and high-level classification. Detecting human figures is formulated as a problem of finding maxima of the distribution generated in the Haar cascade's response space. To find these maxima, we employ an augmented Mean Shift algorithm which assigns each hypothesis a confidence measure related to the distance of the classifier's response from the decision boundary. This confidence is incorporated in the Mean Shift formula to direct the search towards the more reliable hypotheses, which ensures a more accurate detection. On top of the detection framework we have developed a Hidden Markov Model action classifier. It is based on a discrete set of key poses obtained via clustering and utilises a simple motion feature representation. Our approach is exemplified by gait recognition. The model has been trained to recognise four different types, separately in the left and right direction, using the video sequences obtained with a stationary DV camcorder. Good performance on the unseen sequences has been observed which validates our approach and suggests it could be adopted for more general action recognition. Keywords-human detection; gait recognition; mean shift; motion feature representation