Pedestrian detection using GPU-accelerated multiple cue computation

Achieving accurate pedestrian detection for practically relevant scenarios in real-time is an important problem for many applications, while representing a major scientific challenge at the same time. In this paper we present an algorithmic framework which efficiently computes pedestrian-specific shape and motion cues and combines them in a probabilistic manner to infer the location and occlusion status of pedestrians viewed by a stationary camera. The articulated pedestrian shape is represented by a set of sparse contour templates, where fast template matching against image features is carried out using integral images built along oriented scan-lines. The motion cue is obtained by employing a non-parametric background model using the YCbCr color space. Both cues are computed and evaluated on the GPU. Given the probabilistic output from the two cues the spatial configuration of hypothesized human body locations is obtained by an iterative optimization scheme taking into account the depth ordering and occlusion status of individual hypotheses. The method achieves fast computation times even in complex scenarios with a high pedestrian density. Employed computational schemes are described in detail and the validity of the approach is demonstrated on three PETS2009 datasets depicting increasing pedestrian density. Evaluation results and comparison with state of the art are presented.