Laser-Based People Detection and Obstacle Avoidance for a Hospital Transport Robot

Kuisong Zheng, Feng Wu, Xiaoping Chen

TL;DR

This paper presents a novel deep learning approach for robust, real-time people detection using laser scans, coupled with a human-aware obstacle avoidance system for hospital transport robots. By combining efficient point-cloud processing with a sampling-based local path planner, the system aims to improve the safety and naturalness of robot navigation in crowded hospital environments.

What Problem Does This Paper Address?

Hospitals are increasingly adopting autonomous transport robots to improve efficiency and reduce manual labor. However, navigating these robots safely and effectively in dynamic, human-populated environments poses significant challenges. Traditional methods for detecting people using laser data often rely on hand-crafted features and require extensive manual labeling, making them difficult to deploy and scale. Furthermore, purely reactive obstacle avoidance algorithms can lead to jerky, unpredictable movements that confuse pedestrians and compromise safety. This paper tackles these issues by developing a more robust and efficient people detection system and a more human-friendly obstacle avoidance strategy for hospital robots.

Methods

The paper introduces two main contributions:

  1. Laser-Based People Detection:
    • Deep Learning with PointNet: They adapt PointNet, a deep learning architecture designed for point clouds, to classify individual points in a laser scan as either "person" or "non-person." This approach avoids the need for hand-crafted features.
  2. Jump Distance Clustering for Speed: To accelerate inference, they use a "jump distance clustering" method to group laser scan points. This allows the system to evaluate only a subset of points (centroids of clusters) for classification, significantly reducing computational load without sacrificing accuracy.
  3. Automated Dataset Annotation: To address the laborious task of manually labeling laser scan data, they propose a method to automatically annotate datasets. This involves calculating a background model and then identifying points that deviate from it as potential people.
  4. Human-Aware Obstacle Avoidance:
    • Sampling-Based Local Path Planner: Instead of purely reactive methods, they employ a sampling-based local path planner, inspired by autonomous driving systems. This planner generates multiple candidate paths by sampling laterally and longitudinally around the robot's reference path.
  5. Lane Switching for Naturalness: The core idea is to "switch lanes" to avoid obstacles, mimicking human behavior. This approach aims to produce more predictable and less oscillatory movements.
  6. Objective Function for Safety and Legibility: A multi-objective cost function is used to select the best candidate path. This function considers proximity to obstacles, deviation from the reference path, and a "social norm" that encourages passing on the right, promoting socially compliant navigation.
  7. Stop-Wait-Warn Rule: For dynamic situations, a simple but effective rule is adopted: if an obstacle is encountered, the robot stops, warns, and waits for the obstacle to move. This reduces unnecessary deviations and improves efficiency.

Datasets

  • Training and Testing Datasets: The authors collected their own datasets in various environments, including teaching buildings, canteens, and hospital corridors.
  • Dynamic Datasets: Recorded with a robot in sites with significant pedestrian traffic.
  • Static Datasets: Collected by manually controlling the robot in environments devoid of people.
  • Automated Annotation: The proposed automated annotation method was used to label the dynamic datasets. The paper mentions using a threshold of 0.2 for this process.
  • Dataset Split: Four dynamic and three static datasets were used for training, while the remaining were used for testing. The test set was designed to cover different areas than the training set to evaluate generalization.

Key Insights and Ideas

  • Deep Learning for Laser Scans: Applying deep learning directly to raw laser scan data, specifically using PointNet, can effectively replace traditional hand-crafted feature engineering for people detection, leading to better performance.
  • Speed-Accuracy Trade-off Management: The jump distance clustering method is a clever way to significantly speed up inference for point-cloud based methods by reducing the number of points that need to be processed, without a substantial loss in accuracy.
  • Automated Annotation is Crucial: The manual labeling of large datasets is a major bottleneck. The proposed automated annotation method is a significant contribution for enabling the practical deployment of such systems.
  • Human-like Navigation is Key: Purely reactive obstacle avoidance is insufficient for human environments. Mimicking human navigation strategies like lane switching and incorporating social norms leads to more predictable and acceptable robot behavior.
  • Simplicity Can Be Effective: The "stop-wait-warn" rule, while simple, proves effective in many dynamic scenarios, reducing unnecessary complex maneuvers and improving efficiency.

Key Results

  • People Detection Performance: The proposed method achieved comparable performance to state-of-the-art approaches like DROW, with a pixel accuracy (pAcc) of 91.8% and a mean IoU (mIoU) of 85.8%. It outperformed CNN and PointNet on their specific dataset.
  • Inference Speed: The jump distance clustering method accelerated inference by approximately 5x.
  • Obstacle Avoidance: Qualitative experiments demonstrated the effectiveness of the proposed obstacle avoidance algorithm in both static and dynamic scenarios. The robot successfully navigated through static obstacles and handled crossing pedestrians by stopping and waiting, exhibiting predictable and safe behavior.
  • Generalization: The system showed good generalization capabilities to unseen environments, as evidenced by the performance on the test set.

Limitations and Future Work

  • Limitations:
    • The current system cannot autonomously control elevators and doors, requiring human assistance for these tasks.
  • The paper notes that in very narrow spaces (e.g., doors, elevators), situations can arise where neither the robot nor a human can move, leading to a standstill.
  • The proposed people detection method is primarily designed for 2D laser scans, though it's noted that it can be adapted for 3D or multi-layer scanners.
  • Future Work:
    • Adding a "backward" behavior to the robot's state machine to handle situations where giving way is necessary in narrow spaces.
  • Developing more sophisticated prediction of human paths to better handle complex interactions.
  • Making the robot's behavior "smarter" by adjusting waiting times based on obstacle types and choosing lanes based on the orientation and speed of nearby people.
  • Further integration of camera and stereo camera data for richer environmental understanding (currently planned for future work).

Summary updated

Abstract

This paper describes the development of a laser-based people detection and obstacle avoidance algorithm for a differential-drive robot, which is used for transporting materials along a reference path in hospital domains. Detecting humans from laser data is an important functionality for the safety of navigation in the shared workspace with people. Nevertheless, traditional methods normally utilize machine learning techniques on hand-crafted geometrical features extracted from individual clusters. Moreover, the datasets used to train the models are usually small and need to manually label every laser scan, increasing the difficulty and cost of deploying people detection algorithms in new environments. To tackle these problems, (1) we propose a novel deep learning-based method, which uses the deep neural network in a sliding window fashion to effectively classify every single point of a laser scan. (2) To increase the speed of inference without losing performance, we use a jump distance clustering method to decrease the number of points needed to be evaluated. (3) To reduce the workload of labeling data, we also propose an approach to automatically annotate datasets collected in real scenarios. In general, the proposed approach runs in real-time and performs much better than traditional methods. Secondly, conventional pure reactive obstacle avoidance algorithms can produce inefficient and oscillatory behaviors in dynamic environments, making pedestrians confused and possibly leading to dangerous reactions. To improve the legibility and naturalness of obstacle avoidance in human crowded environments, we introduce a sampling-based local path planner, similar to the method used in autonomous driving cars. The key idea is to avoid obstacles by switching lanes. We also adopt a simple rule to decrease the number of unnecessary deviations from the reference path. Experiments carried out in real-world environments confirmed the effectiveness of the proposed algorithms.
BibTeX
@inproceedings{zheng2021laser,
  author = {Kuisong Zheng and Feng Wu and Xiaoping Chen},
  title = {{Laser-Based People Detection and Obstacle Avoidance for a Hospital Transport Robot}},
  booktitle = {Sensors},
  year = {2021},
  url = {https://openreview.net/forum?id=HBwx9lMkiEG},
}
Figure