Pedestrian Collision Danger Model using Attention and Location Context

Intelligent and autonomous vehicle safety is a rapidly developing field. With the increasing number of electric vehicles as well as following consumer trends, cars are getting quieter and also heavier which may lead to severe traffic accidents. To help avoiding potential dangerous situations leading to accidents, this paper proposes a collision danger model for individual pedestrians that can aid vehicle safety features and help decision making, using only forward facing optical cameras. Multi pedestrian detection and tracking is performed with a fast joint model. Semantic segmentation and classification is used to refine pedestrian contours and find the 3D positions as well as to understand the location context of pedestrians in the environment. Pedestrian position is tracked and orientation is estimated using 2D bounding boxes. The proposed pedestrian danger model is the combination of the awareness estimated from orientation, passing distance estimated from trajectories and location context from the segmentation results.