PDCS10

Amongst the various capabilities required by an intelligent vehicle perception system, object detection is a key one to provide safe and reliable vehicle autonomy or assistance (e.g. collision detection). Over the plethora of world objects which a human being learns how to avoid in driving situations, the human being himself is one of the most difficult for an intelligent machine to detect. This is so because the human body can appear in several poses, positions, forms, sizes and colors. Research in the area of pedestrian detection has been intense in the last years in our research group with state of the art results. However there are some 'critical' situations, regarding pedestrian detection on urban scenarios that were not satisfactory resolved. The key problem behind these critical situations occurs when pedestrians appear in the scene partially occluded, close to each other (cluttered situations) or in scenes with complex backgrounds (problematical object-background segmentation). New challenges are now faced by the research team on developing pedestrian detection systems suitable for more real-world applications, focused on specific challenging urban scenarios. These new approaches will embody scene contextual information aiming to improve the detection performance in the above mentioned challenging situations, while performing also the assessment of danger situations. To accomplish the project objectives, the following tasks were defined: 1) Laser and vision data segmentation; 2) Feature Extraction and Feature Selection; 3) Neural Classifiers - Training; 4) Classifier fusion - Trainable approaches; 5) Context-aware multi-sensor fusion; 6) Scenarios definition, datasets and field tests.