HOG based multi-object detection for urban navigation

A necessary condition to perform a fully autonomous driving system in urban environment is to detect object types in real scenes. Visual object recognition is a key solution, but multi-object detection still remain unsolved. In this paper, we present a fast and efficient multi-object detection system built to recognize, at the same time, pedestrians cars and bicycles. For each target type, we construct a holistic detector in a cascade manner, using a dense overlapping grid based on histograms of oriented gradients (HOG). The selection of HOG features is obtained through a learning process using AdaBoost algorithm. Experiments have been conducted on the car-like robot Robucar, where the single detectors are combined and implemented on its embedded computer, which is endowed with a modular software platform. Results are promising as the system can process up to 20 fps with VGA images.