Quasi-Random Sampling for Condensation

Abstract

Abstract. The problem of tracking pedestrians from a moving car is a challenging one. The Condensation tracking algorithm is appealing for its generality and potential for real-time implementation. However, the conventional Condensation tracker is known to have di culty with high-dimensional state spaces and unknown motion models. This paper presents an improved algorithm that addresses these problems by using a simplified motion model, and employing quasi-Monte Carlo techniques to e ciently sample the resulting tracking problem in the high-dimensional state space. For N sample points, these techniques achieve sampling errors of O(N 31 ), as opposed to O(N 31/2 ) for conventional Monte Carlo techniques. We illustrate the algorithm by tracking objects in both synthetic and real sequences, and show that it achieves reliable tracking and significant speed-ups over conventional Monte Carlo techniques.

Monocular Pedestrian Detection: Survey and Experiments

“…One line of research has formulated tracking as frame-by-frame association of detections based on geometry and dynamics without particular pedestrian appearance models [2], [23]. Other approaches utilize pedestrian appearance models (Section 2.2) coupled with geometry and dynamics [4], [26], [32], [39], [43], [50], [55], [58], [65], [70], [76], [77], [80], [82]. Some approaches furthermore integrate detection and tracking in a Bayesian framework, combining appearance models with an observation density, dynamics, and probabilistic inference of the posterior state density.…”

“…Some approaches furthermore integrate detection and tracking in a Bayesian framework, combining appearance models with an observation density, dynamics, and probabilistic inference of the posterior state density. For this, either single [4], [26], [55], [70], [76] or multiple cues [32], [43], [50], [58], [65] are used.…”

Multi-cue Pedestrian Detection and Tracking from a Moving Vehicle

“…Representations such as Active Shape Models (Baumberg and Hogg, 1994;Cootes et al, 1995;Philomin et al, 2000), shape exemplars (Stenger et al, 2003;Toyama and Blake, 2001) and color blobs (Heisele and Wöhler, 1998) have been combined with Kalman- (Baumberg and Hogg, 1994;Cootes et al, 1995) or particle filtering (Philomin et al, 2000;Stenger et al, 2003;Toyama and Blake, 2001) approaches. Given temporal ROI data, previous work has detected people walking laterally to the viewing direction, either using the periodicity cue (Cutler and Davis, 2000;Polana and Nelson, 1994) or by learning the characteristic lateral gait pattern (Heisele and Wöhler, 1998).…”

Pedestrian tracking from a moving vehicle

“…Even when one uses a "perfect" pseudo-random sequence for generating N sample points, the sampling error will only decrease as O(N −1/2 ) as opposed to O(N −1 ) for another class of sequences known as quasi-random sequences which have low discrepancy. We introduced quasi-random sampling in the context of the Condensation algorithm in Philomin et al [10] and showed that even in low dimensions, a significantly fewer amount of sample points were needed to achieve the same sampling error when compared to pseudo-random sampling. For reasons of brevity, the details are not discussed here; the readers are referred to [10].…”

“…We introduced quasi-random sampling in the context of the Condensation algorithm in Philomin et al [10] and showed that even in low dimensions, a significantly fewer amount of sample points were needed to achieve the same sampling error when compared to pseudo-random sampling. For reasons of brevity, the details are not discussed here; the readers are referred to [10]. In typical implementations of the Condensation algorithm, a "perfect" pseudorandom number generator is almost never used and a linear congruential generator (such as the system supplied rand()function) is used instead.…”