Spatio-Temporal Perception for Mobile Robots in Dynamic Environments

Robotic systems have the potential to revolutionize operations in environments that are too dangerous, intricate, or demanding for humans. These environments pose notable challenges due to their dynamic and often unpredictable nature, requiring robots to identify and adapt to short- and long-term changes. By leveraging advanced perception capabilities, robots can address critical tasks such as preventing accidents by detecting and reacting to vulnerable road users. They can also autonomously transport humans and goods while adapting to evolving demands, carrying out time-consuming tasks like crop monitoring, or accomplishing dangerous missions like disaster response. Robots can execute tasks more efficiently and safely by overcoming human limitations such as fatigue, inattention, or restricted sensory perception. <br> In these scenarios, mobile robots typically operate autonomously, continuously perceiving their environment to estimate both their internal state and the state of their surroundings. They usually rely on sensors like global navigation satellite system receivers, cameras, radar sensors, inertial measurement units, or LiDAR scanners. Key tasks include building maps of the environment, localizing in such maps, or segmenting different classes like cars, buildings, or traffic signs. Urban environments are often complex and dynamic, containing moving objects like humans or undergoing structural changes. To solve such tasks, a mobile robot must possess both spatial awareness and spatio-temporal perception - an understanding of how the environment evolves and what is changing in it explicitly. A major challenge these approaches encounter is the unknown nature of the environment beforehand, which requires the system to be highly robust and able to generalize across different sensor configurations and settings. <br> This thesis focuses on two main questions when deploying mobile robots in unknown and dynamic environments: “What is moving?” and “Where is an object moving to?”. We must process and interpret spatial and temporal data to address these. First, knowing which parts of the environment belong to moving objects is an essential spatio-temporal perception task for online path planning. For example, moving objects occupy space only temporarily, meaning we can consider the space again traversable for planning after the object has moved. Moving objects can also advance into areas previously regarded as free, causing potential collisions with our planned trajectory. Second, we are interested in estimating the future state of the surroundings. This prediction enables us to, for example, properly plan a future path that reflects the future behavior of other traffic participants.