Monocular detection of pedestrians for the safe indoor navigation of a mobile robot

When a mobile robot is employed in an unstructured space, safety during its motion should be considered first. In particular, if the robot shares a workspace with humans, detecting humans in front of the navigating robot is of great importance in order to avoid collision. Although low-cost sensors based on ultrasound or infrared are widely used to avoid the robot´s collision, they have limitations of low resolution and low detecting range, and this makes difficult for the robot to do human-like efficient obstacle avoidance. In this paper, we present a framework for detecting pedestrians using a single camera mounted on a robot that moves in a corridor. The body detector using Histogram of Oriented Gradients method and the face detector using AdaBoost are used in a complementary manner. A detected pedestrian is approximated by a bounding box and his/her position on the floor is estimated from the low end of the box referencing the left and right boundaries of the floor in the image, which are detected by Hough Transform. The estimated human positions are tracked in consecutive images so as for the robot to decide a proper action. The proposed method was verified in experiments using real indoor video images.