Intelligent Pedestrian Intention Prediction Framework

Abstract

One of the most critical tasks in autonomous driving is anticipating pedestrian crossing intention on roads to ensure safe and reliable driving. This will instil trust in the road user community in driving assistance endeavours from Advanced Driving Assistance Systems (ADAS) to Autonomous Vehicles (AVs) encouraging their co-existence. In this paper, a cascade of three modules is employed, the convolution module that acts as a feature extractor, the recurrent module that is used for sequential tasks followed by classification module. It is shown that with the help of information regarding the past trajectory, appearance of pedestrians and the ego-vehicle speed, the proposed data-driven approach is able to predict pedestrian crossing intention reliably. The proposed algorithm is able to anticipate crossing intention in two publicly available benchmark datasets, JAAD and PIE with an accuracy of 88% and 86% respectively.