Human Pose Estimation in Real Traffic Scenes

Abstract

This work evaluates human pose estimation from a moving vehicle for use in road traffic applications such as automated or autonomous driving. The lack of annotated human pose datasets is a challenge for research on autonomous driving with sufficient safety for pedestrians and cyclists. Hence, a dedicated dataset was created, which was recorded in real traffic and contains both pedestrians and cyclists. The dataset represents diverse conditions in road traffic and different appearances of people and allows for a realistic evaluation of the performance of human pose estimation. In order to obtain ground truth, the data were labeled manually and $3D$ poses were measured by means of an intersection equipped with a wide angle stereo camera system. One recent method for $2D$ pose estimation and another approach for $3D$ pose estimation from literature are investigated. As shown by this research, $2D$ poses estimated in traffic scenes achieve a similar accuracy as state-of-the-art results obtained on other datasets. Moreover, $3D$ pose estimation based on single images outperforms a naive distance measurement of individual joints in disparity maps obtained by a stereo camera and the reached accuracy suggests that the use of $3D$ pose estimation for intention detection and trajectory forecast is feasible. Overall, a dependency of the results on the distance of the respective person and less reliable results for cyclists compared to pedestrians were observed.