A pedestrian detection module for the USAD vehicle

Return to USAD, main project page

Presentation of my thesis - Presentazione Riccardo Capra (17 downloads)

The goal of this work is to build a sistem able to detect and locate the pedestrians in front of the vehicle.
In order to do that we use the two stereoscopic cameras mounted on the USAD vechicle to get a point cloud representing the 3D space in front of the vehicle.
A neural network (Yolo v3) makes the detections on the frames from the master camera returing the bounding boxes.
We merge the 2D information from the neural network with the 3D point cloud and, scanning the 3D points that correspond to the 2D point of the detected people, we are able to estimate the people's distance from the front of the vehicle.
In the end we publish the information relative to the distances of all the people and the points used to calculate them; the node responsible for the dynamic obstacle management will use this information to avoid those people on the road.