Experimental Evaluation of Multi-cue Monocular Pedestrian Detection System Using Built-In Rear View Camera

Abstract

Algorithms for pedestrian detection based on multi-cue computer vision systems have become increasingly sophisticated and have been shown capable of achieving high detection performance for moderate range of applications. In particular, it has been demonstrated that a key successes lies in the integration of both single frame and over time measured cues and via building up additional object categories consisting of vehicles and stationary background structures. However, most of extensive field tests were made with a use of high contrast, good resolution, large filed of view, distortion free mounted cameras. In this paper we use a standard vehicle with a built-in rear view camera for parking space detection that has moderate contrast, low resolution, limited field of view and large distortions. For the evaluation of multi-cue pedestrian recognition system, we used a database of more than 300 typical sequences from 30 seconds to 30 minutes of 30 km/h limited drive both in Europe and Japan. Using that database and semi-automatically collected ground truth, several experiments were carried out to determine how pedestrian detection performance evolves with respect to distance from vehicle to pedestrian, vehicle's speed and illumination conditions. The results and interpretation of these experiments will be described in this paper.