Driver attention Assistance by Pedestrian/cyclist distance estimation from a single RGB Image: A CNN-based semantic segmentation approach

Nowadays, automotive companies are investing a relevant amount of resources for designing autonomous driving systems, driver assistance technologies and systems for assessing the driver’s attention. In this context, an important application consists of technologies for estimating the object distances in the scene, with a specific focus on pedestrians/cyclists. These technologies are usually based on LiDAR scanners, and thus require dedicated sensors and post-processing algorithms for estimating a depth map representing the distances between the vehicle and the surrounding objects. To obtain highly accurate distance estimations, methods based on Deep Learning (DL) and Convolutional Neural Networks (CNN) are being increasingly used for semantic segmentation in autonomous driving applications, considering either RGB images or LiDAR scans. In this paper, we propose the first method in the literature able to estimate the distances of pedestrians from the vehicle by using only an RGB image and CNNs, without the need for any LiDAR scanner or any device designed for three-dimensional reconstruction of the scene. The proposed method is based on two CNNs: the first one semantically segments the image regions representing pedestrians/cyclists, while the second one (DepthCNN) estimates a dense depth map of the scene. We evaluated our approach on a public dataset of RGB images and LiDAR scans captured in an automotive scenario, with results confirming the feasibility of the proposed method.