Keywords: robotic vision, mapping, navigation, machine learning, deep learning
While moving in an environment, we intuitively avoid colliding with other people by following implicit social norms and predicting each other’s trajectories from movement cues. Socially compliant navigation is a crucial skill for robots to be employed in human-inhabited environments. For indoor robots, maps are used as internal representation of the environment where a robot would be navigating. These maps encode different representations of the world — e.g. obstacle occupancy or object semantics — and are built autonomously by the robot while moving and observing the environment through different sensors.
This thesis goal is to develop a probabilistic approach for building and updating people flow maps able to help the robot being socially-aware while navigating by predicting people movement. These maps can be initialized from a Deep Learning prior, and then updated using statistical inference from real evidence of people moving around the robot.
- Review of relevant state-of-the-art literature
- Design and development of a probabilistic approach to people flow mapping
- Integration of state-of-the-art machine learning approaches for people movement mapping and people tracking with the proposed method.
- Evaluation of system performance in real life scenarios and on real robots
Prerequisites: C++/python (high), deep learning (medium), probabilistic methods (medium), Linux (medium)
Tools: pyTorch/tensorflow, ROS, Care-O-bot 4
Start: Available immediately