Layered Path Planning with Human Motion Detection for Autonomous Robots

Huan Tan, Yang Zhao, Lynn A. DeRoseInTech 2017 · DOI · OpenAlex

Abstract

Reactively planning a path in a dynamic and unstructured environment is a key challenge for mobile robots and autonomous systems. Planning should consider factors including the long-term and short-term prediction, current environmental situation, and human context. In this chapter, we present a novel robotic path-planning method with human activity information in a large-scale three-dimensional (3D) environment. In the learning stage, this method uses human motion detection results and preliminary environmental information to build a long-term context model with a hidden Markov model (HMM) to describe and predict human activities in the environment. In the application stage, when a robot detects humans in the environment, it first uses the long-term context model to generate impedance areas in the environment. Then, the robot searches each area of the environment to find paths between key locations, such as escalators, to generate a Reactive Key Cost Map (RKCM), whose vertexes are those key locations and edges are generated paths. The graphs of all areas are connected using the key nodes in the subgraphs to build a global graph of the whole environment. Finally, the robot can reactively plan a path based on the current environmental situation and predicted human activities. This method enables robots to navigate robustly in a large-scale 3D environment with regular human activities, and it significantly reduces computing workload with proposed RKCM.

Introduction

Navigating complex, dynamic environments like hospitals or office buildings remains a significant challenge for autonomous mobile robots. Traditional navigation systems often react to obstacles only when they are within the immediate sensing range, leading to inefficient paths or safety concerns in crowded areas. To address this, robots must not only detect current obstacles but also anticipate human activity patterns. This paper proposes a layered path-planning framework that integrates long-term human activity prediction with a computationally efficient multi-floor navigation strategy.

System Architecture

The proposed system operates in two distinct phases: a learning stage and an application stage. The architecture is designed to bridge the gap between high-level global planning and low-level reactive avoidance by incorporating human context into the cost maps used for pathfinding.

The proposed system architecture integrates human activity detection and long-term context modeling into a Reactive Key Cost Map for global path generation.

As shown in the architecture diagram, the process begins with human activity detection, which feeds into a long-term context model. This model, combined with the specific navigation task, informs the creation of a Reactive Key Cost Map (RKCM). The RKCM is then integrated into a global map to generate the final optimal path.

Human Activity Modeling and Sensing

A core component of this approach is device-free human tracking using Radio Tomographic Imaging (RTI). By deploying a network of wireless nodes, the system measures signal attenuation caused by human bodies to create occupancy heat maps.

Heat maps generated from Radio Tomographic Imaging (RTI) identify 'hot spots' of human activity, which are used to build the long-term context model.

These heat maps are processed using Gaussian Mixture Models (GMMs) to identify "hot spots"—areas with high frequencies of human presence. A Hidden Markov Model (HMM) is then employed to model the temporal transitions between these spots, allowing the robot to predict where crowds are likely to form at different times of the day.

Reactive Key Cost Map (RKCM)

To handle the complexity of large-scale 3D environments, the researchers introduced the Reactive Key Cost Map. Instead of planning on a high-resolution 3D grid, the environment is abstracted into a graph of "key locations" such as elevators, stairs, and destination points. The edges between these nodes are weighted based on the predicted human activity levels derived from the HMM. This layered approach significantly reduces the computational overhead, allowing the robot to find global minimum-cost paths across multiple floors efficiently.

Results and Discussion

The framework was validated through both physical experiments and simulations. In a residential testbed with 34 wireless nodes, the RTI system accurately tracked human movement and identified peak activity zones. In multi-floor hospital simulations, the RKCM-based planner successfully navigated the robot through the least crowded routes, accounting for the costs associated with vertical transitions (elevators) and high-traffic areas.

The results demonstrate that by treating crowded regions as "impedance areas" in advance, the robot can proactively avoid bottlenecks. This reduces the need for frequent, computationally expensive local re-planning and improves the overall flow of robotic operations in human-centric environments.

Conclusion

This research presents a robust method for robotic navigation that respects human context. By combining device-free sensing, probabilistic activity modeling, and an abstracted graph-based planning approach, robots can navigate large-scale 3D environments more intelligently and efficiently. Future work will focus on lifelong learning to adapt to evolving human patterns and extending the framework to multi-robot coordination.

Summary updated

Abstract

Reactively planning a path in a dynamic and unstructured environment is a key challenge for mobile robots and autonomous systems. Planning should consider factors including the long-term and short-term prediction, current environmental situation, and human context. In this chapter, we present a novel robotic path-planning method with human activity information in a large-scale three-dimensional (3D) environment. In the learning stage, this method uses human motion detection results and preliminary environmental information to build a long-term context model with a hidden Markov model (HMM) to describe and predict human activities in the environment. In the application stage, when a robot detects humans in the environment, it first uses the long-term context model to generate impedance areas in the environment. Then, the robot searches each area of the environment to find paths between key locations, such as escalators, to generate a Reactive Key Cost Map (RKCM), whose vertexes are those key locations and edges are generated paths. The graphs of all areas are connected using the key nodes in the subgraphs to build a global graph of the whole environment. Finally, the robot can reactively plan a path based on the current environmental situation and predicted human activities. This method enables robots to navigate robustly in a large-scale 3D environment with regular human activities, and it significantly reduces computing workload with proposed RKCM.
BibTeX
@inproceedings{tan2017layered,
  author = {Huan Tan and Yang Zhao and Lynn A. DeRose},
  title = {{Layered Path Planning with Human Motion Detection for Autonomous Robots}},
  booktitle = {InTech},
  year = {2017},
  doi = {10.5772/intechopen.68145},
}
Figure