A Real-time LIDAR and Vision based Pedestrian Detection System for Unmanned Ground Vehicles

In this work, we present a real-time pedestrian detection system using LIDAR and Vision in-vehicle. We get regions of interest by clustering lidar point clouds and project them onto the images. After that we use black mask to replace those image areas which has no lidar points projected onto. Then we extract HOG and lidar point clouds features and use those features to detect pedestrians by a linear SVM classifier. The main contributions are that we proposed a method that can select ROIs on image automatically and then enhanced the HOG descriptor with the lidar points' projections. Finally we fuse HOG and lidar based features to train a linear SVM to detect pedestrian. The above method we proposed can satisfy real-time requirement. We apply our pedestrian detection system to our own dataset and KITTI dataset, and show that we outperform the primitive HOG based methods.