I'm Walking Here!: Pedestrian Intent Recognition Identifying Future Pedestrian Trajectory using Machine Learning On-Board an Autonomous Vehicle

Abstract

For autonomous cars to become street viable, they must go beyond sensing pedestrians to predicting the actions of pedestrians and their future positions. Pedestrian Intent Recognition (PIR) is a system for predicting future positions and trajectories of pedestrians on board an autonomous vehicle. Using the OpenPose library as a basis, a model was created using a feed forward neural network based on 3D skeletal posture to predict future pedestrian positions. The predictive model results in an accurate depiction of pedestrian actions as trained on the Carnegie Mellon Mo-Cap Dataset. PIR functions as a proof of concept that can be adapted to an autonomous vehicle platform. Future work includes adapting the model to function using LIDAR and multiple RGB cameras as input for a real-time prediction system for autonomous vehicles.