Pedestrian Motion Prediction Evaluation for Urban Autonomous Driving

Pedestrian motion prediction is a key aspect in any autonomous-driving pipeline, and is required for ensuring safe, accurate, and timely awareness of human agents' possible future trajectories. Autonomous vehicles need to use agent motionprediction information to prevent any possible accidents, and for creating a comfortable and pleasant driving experience for vehicles' passengers as well for pedestrians in vehicles' vicinity. A significant amount of research has been conducted on the topic of agent motion prediction in the fields of robotics, computer vision, and intelligent transportation systems etc. However, a relatively unexplored aspect in the existing literature is the integration of state-of-the-art motion-prediction solutions into existing autonomous driving stacks and evaluating them in real-life conditions rather than on sanitized datasets. In this paper, we analyze a set of selected methods from the literature, and present the perspective obtained by integrating them into an existing autonomous-driving software stack - Autoware Mini - and performing experiments in natural urban conditions in Tartu, Estonia to determine the suitability of conventional motion prediction metrics. Our study should be of value to researchers in autonomous driving or robotics interested in understanding real-world performance of existing state-of-theart pedestrian motion prediction methods. The code, along with instructions on accessing the dataset that we employ, is available at https://github.com/dmytrozabolotnii/autoware_mini