Exploring Sensor Fusion Schemes for Pedestrian Detection in Urban Scenarios

This work explores three schemes for pedestrian detection in urban scenarios using information gathered by a LIDAR and a monocular camera mounted on an electric vehicle. In the first scheme, pedestrian detection is conduct by a set of single classification methods trained with LIDAR and/or vision-based features. In the second scheme, the likelihoods from the single-classifiers are fused by means of three fusion rules: average, maximum value and a Naive-product rule. Finally, the third scheme is a cascade of classifiers with fourstages, two per feature space. All these pedestrian detection strategies were compared through off-line experiments conducted using our dataset, that is available on the Web for public usage.