Efficient combination of Lidar intensity and 3D information by DNN for pedestrian recognition with high and low density 3D sensor

Abstract

Pedestrian recognition is one of the key components for assisted and autonomous driving. So far many researchers have investigated systems combining a high density LIDAR with cameras or stereo, which results in an expensive and complex setup where the LIDAR data is mostly used to extract regions of interest for the 2D sensor. Very few work has focused on using pure 3D data coming from the LIDAR to recognize pedestrians, and even less have made an intensive use of the intensity information returned by the LIDAR. The intensity information displays a high frequency change between neighboring points of similar material, this can be due to the angle or distance. Due to this, it has not been frequently investigated as a potentially interesting feature as it would require extensive time consuming feature engineering to be worthwhile. In this paper we present a novel 2D representation of a 3D point cloud including the intensity information. We show the ability of convolutional neural networks to handle this data in order to accurately recognize pedestrians in complex driving scenes. Our system outperformed state of the art technique on the STC database. Additionally we show that this system is still highly accurate on low density LIDAR data.