Pedestrian Intention Recognition and Path Prediction for Video-based Advanced Driver Assistance Systems
Shekhar Raheja
Masters Thesis

Abstract:

This work presents a novel approach for video-based advanced driver assistance systems to predict an approaching pedestrian's intention to cross the road or not and accordingly estimate the future path. The presented approach addresses the most frequent real world problem faced by drivers to correctly estimate if the approaching pedestrian would cross the road or not. The primary goal of this research is to reduce the road accident casualties by automatically bringing the vehicle to a halt upon seeing an approaching pedestrian in high risk zone in case the driver is unaware. Features consisting of motion information in 3D-space and head-pose orientation over time are fed into the system. The system performs intention prediction based on time-series analysis and probabilistic trajectory matching within a approximated particle filtering-like framework. The system further estimates the pedestrian's path one second ahead in the future and uses the information from intention recognition system to improve path prediction.

Keywords:

Pedestrian Intention Recognition, Pedestrian Path Prediction, Porbabilistic Trajectory Matching, Advanced Driver Assistance Systems (ADAS), Snippets, time Series, Head Pose, Binary Search Tree (BST), Stopping Probability

Pedestrian Intention Recognition and Path Prediction for Video-based Advanced Driver Assistance Systems
(Hrsg.)
Masters Thesis

Abstract:

This work presents a novel approach for video-based advanced driver assistance systems to predict an approaching pedestrian's intention to cross the road or not and accordingly estimate the future path. The presented approach addresses the most frequent real world problem faced by drivers to correctly estimate if the approaching pedestrian would cross the road or not. The primary goal of this research is to reduce the road accident casualties by automatically bringing the vehicle to a halt upon seeing an approaching pedestrian in high risk zone in case the driver is unaware. Features consisting of motion information in 3D-space and head-pose orientation over time are fed into the system. The system performs intention prediction based on time-series analysis and probabilistic trajectory matching within a approximated particle filtering-like framework. The system further estimates the pedestrian's path one second ahead in the future and uses the information from intention recognition system to improve path prediction.

Keywords:

Pedestrian Intention Recognition, Pedestrian Path Prediction, Porbabilistic Trajectory Matching, Advanced Driver Assistance Systems (ADAS), Snippets, time Series, Head Pose, Binary Search Tree (BST), Stopping Probability