Human Motion Trajectory Prediction for Socially Compliant Robot Navigation
Basic Information
Students: Markus Himanen, Harsha Guda, Abdelrahman Abdeldaim, Madis Lemsalu
Project manager: Markus Himanen
Instructor: Mohammadreza Nakhaei
Other advisors:
Starting date: 10.1.2023
Completion date: 6.6.2023
Abstract
With the increasing use of autonomous robots in human environments, the ability of these robots to understand and predict human behaviour is crucial for safe and seamless smooth operation. To avoid collisions and disturbances, robots must be able to anticipate human movements and respond accordingly. In this project, we developed an algorithm for mobile robots, specifically the Boston Dynamics Spot robot, to predict future human paths and proactively avoid them.
While Spot has a built-in collision avoidance system for immediate obstacles, however, it lacks the capability to foresee and avoid situations that could lead to a collision. Therefore, it has limited capablities to avoid dynamic obstacles including humans. Our algorithm detects humans using Spot's sensors, estimates their state, and predicts their most probable future trajectories, which we term "danger zones." Spot can then navigate to avoid entering these zones. Such predictive capability could enable socially compliant robot navigation.
We utilized Spot's 3D LiDAR and cameras providing 360° coverage. From this data we detected and localized multiple humans using YOLOV8 and StrongSORT tracking algorithm. We then generated danger zones representing the areas each human was most likely to move into. Our system operated on recordings of Spot's sensor data from ROS.
Presentation
Poster
Demo
The red zone around the people is tracked to represent a zone which robot has to avoid.
