ADBSCAN Follow-me#

ADBSCAN (Adaptive Density-Based Spatial Clustering of Applications with Noise) is an Intel® patented unsupervised clustering algorithm designed for robust spatial object detection and localization from 2D LiDAR, 3D LiDAR, and Intel® RealSense™ depth camera point clouds.

Unlike traditional DBSCAN—which relies on static neighborhood search radii ($\epsilon$) and fixed density thresholds ($MinPts$)—ADBSCAN dynamically scales these parameters based on range from the sensor and the field-of-view point distribution. This compensates for optical beam divergence and spatial point cloud sparsity at extended distances, yielding a 20–30% increase in effective object detection range.

The Follow-Me reference application builds on ADBSCAN to continuously track a target human or guide vehicle, evaluate multi-modal interaction cues (hand gestures and voice commands via Intel® OpwebenVINO™), and command mobile base velocities.

Architecture & Algorithm#

ADBSCAN Algorithmic Foundation#

In LiDAR and structured-light depth sensing, point density decreases non-linearly with distance $r$. Standard spatial clustering with a fixed radius $\epsilon$ causes over-segmentation at close range and cluster fragmentation or omission at distant ranges.

ADBSCAN formulates adaptive clustering parameters:

Dynamic Radius $\epsilon(r)$: Expands monotonically with range to encapsulate sparsely distributed returns from distant objects.
Dynamic Density $MinPts(r)$: Scaled using calibration coefficients (base, coeff_1, coeff_2, scale_factor) to match expected return point distributions at given range slices.
Hardware Acceleration: The neighbor search phase can optionally be offloaded to integrated or discrete GPUs via Intel® oneAPI™ SYCL kernels (oneapi_kdtree or oneapi_octree), delivering substantial latency reductions on dense point clouds.

System Architecture#

The following diagram illustrates the data flow from physical or simulated sensors through perception, interaction, and mobile base control:

        graph TD
    subgraph Sensors["Sensors & Modalities"]
        Lidar["2D / 3D LiDAR<br/>(LaserScan / PointCloud2)"]
        Depth["Intel RealSense Depth Camera<br/>(PointCloud2)"]
        RGB["RGB Camera Stream"]
        Mic["Microphone Audio"]
    end

    subgraph Perception["Perception Layer"]
        ADBSCAN["adbscan_ros2 Node<br/>• Range-Adaptive Radius ε(r)<br/>• Density Scaling MinPts(r)<br/>• Optional oneAPI GPU Offload"]
    end

    subgraph Interaction["Interaction Layer (OpenVINO)"]
        Gesture["gesture_recognition_pkg<br/>(MediaPipe / OpenVINO)"]
        Speech["speech_recognition_pkg<br/>(OpenVINO Speech ASR)"]
        TTS["text_to_speech_pkg<br/>(Synthesized Audio Prompts)"]
    end

    subgraph Interconnect["ROS 2 Topics & Interfaces"]
        Obs["/obstacle_array<br/>(nav2_dynamic_msgs/ObstacleArray)"]
        FMI["follow_me_interfaces<br/>(Gesture / Audio Signals)"]
    end

    subgraph Application["Application Layer"]
        FollowMe["adbscan_ros2_follow_me Node<br/>• Target Cluster Tracking<br/>• Multi-Modal State Machine<br/>• Twist Velocity Generator"]
    end

    subgraph Execution["Actuation & Simulation Targets"]
        CmdVel["/cmd_vel<br/>(geometry_msgs/Twist)"]
        Gazebo["Gazebo Simulation<br/>(TurtleBot3 Waffle + Guide Robot)"]
        Robot["Physical AMR<br/>(Differential Drive Base)"]
    end

    Lidar --> ADBSCAN
    Depth --> ADBSCAN
    RGB --> Gesture
    Mic --> Speech

    ADBSCAN --> Obs
    Obs --> FollowMe

    Gesture --> FMI
    Speech --> FMI
    FMI --> FollowMe

    FollowMe -. Voice Feedback .-> TTS
    FollowMe --> CmdVel
    CmdVel --> Gazebo
    CmdVel --> Robot

Source Code & Workspace#

The component packages are consolidated under the src/ directory:

adbscan_ros2: Core ADBSCAN clustering perception node publishing /obstacle_array.
adbscan_ros2_follow_me: Follow-me tracker, state machine, and velocity generator.
follow_me_interfaces: Custom ROS 2 message and service definitions.
gesture_recognition_pkg: Hand gesture recognition using OpenVINO / MediaPipe.
speech_recognition_pkg: Voice command recognition using OpenVINO ASR models.
text_to_speech_pkg: Audio prompt generation and feedback.
followme_turtlebot3_gazebo: Gazebo simulation environments (Harmonic / Fortress).

Deploy Follow-me on a Physical Mobile Robot (Clearpath Jackal / AAEON AMR)#

This section provides instructions for running the ADBSCAN-based Follow-me pipeline on physical Autonomous Mobile Robots using RealSense depth camera or LiDAR inputs.

The sensor publishes point clouds to /camera/depth/color/points (RealSense) or laser scans to /scan (LiDAR). The adbscan_sub_node or adbscan_sub_w_gesture_audio subscribes to the corresponding topic, segments clusters using range-adaptive parameters, determines the target person’s centroid, computes linear and angular velocity adjustments, and publishes differential-drive commands to /cmd_vel of type geometry_msgs/msg/Twist. This message is subsequently consumed by the chassis base controller.

Deployment Setup#

Deployment Prerequisites#

Complete the Getting Started guide before continuing.

Install the Deployment Packages#

Deployments can run from the prebuilt Debian packages installed from the Robotics AI Dev Kit APT repository:

Jazzy

sudo apt update
sudo apt install ros-jazzy-adbscan-ros2 ros-jazzy-adbscan-ros2-follow-me ros-jazzy-follow-me-interfaces

Humble

sudo apt update
sudo apt install ros-humble-adbscan-ros2 ros-humble-adbscan-ros2-follow-me ros-humble-follow-me-interfaces

Alternatively, if building directly from source in a cloned workspace:

make build
source install/setup.bash

Run Demo#

To launch the follow-me application on a robot platform using Intel RealSense depth sensing, run:

Jazzy

source /opt/ros/jazzy/setup.bash
# If running from a built workspace:
# source install/setup.bash

ros2 launch adbscan_ros2_follow_me play_demo_realsense_launch.py

Humble

source /opt/ros/humble/setup.bash
# If running from a built workspace:
# source install/setup.bash

ros2 launch adbscan_ros2_follow_me play_demo_realsense_launch.py

For 2D LiDAR based tracking:

Jazzy

source /opt/ros/jazzy/setup.bash
ros2 launch adbscan_ros2_follow_me play_demo_lidar_launch.py

Humble

source /opt/ros/humble/setup.bash
ros2 launch adbscan_ros2_follow_me play_demo_lidar_launch.py

After starting the node, the robot begins searching for trackable clusters in its initial detection radius (init_tgt_loc, default 0.5m–0.8m) and tracks the centroid of the acquired target as it moves.

Configuration parameters can be inspected and modified in: /opt/ros/<distro>/share/adbscan_ros2_follow_me/config/adbscan_sub_RS.yaml (or src/adbscan_ros2_follow_me/config/adbscan_sub_RS.yaml when developing locally). Find a brief description of the parameters in the following list:

Lidar_type

Type of the point cloud sensor. For RealSense camera and LIDAR inputs, the default value is set to RS and 2D, respectively.
Lidar_topic

Name of the topic publishing point cloud data.
Verbose

If this flag is set to True, the locations of the detected target objects will be printed as the screen log.
subsample_ratio

This is the downsampling rate of the original point cloud data. Default value = 15 (i.e. every 15-th data in the original point cloud is sampled and passed to the core BSCAN algorithm).
x_filter_back

Point cloud data with x-coordinate > x_filter_back are filtered out (positive x direction lies in front of the robot).
y_filter_left, y_filter_right

Point cloud data with y-coordinate > y_filter_left and y-coordinate < y_filter_right are filtered out (positive y-direction is to the left of robot and vice versa).
z_filter

Point cloud data with z-coordinate < z_filter will be filtered out. This option will be ignored in case of 2D Lidar.
Z_based_ground_removal

Filtering in the z-direction will be applied only if this value is non-zero. This option will be ignored in case of 2D Lidar.
base, coeff_1, coeff_2, scale_factor

These are the coefficients used to calculate adaptive parameters of the ADBSCAN algorithm. These values are pre-computed and recommended to keep unchanged.
init_tgt_loc

This value describes the initial target location. The person needs to be at a distance of init_tgt_loc in front of the robot to initiate the motor.
max_dist

This is the maximum distance that the robot can follow. If the person moves at a distance > max_dist, the robot will stop following.
min_dist

This value describes the safe distance the robot will always maintain with the target person. If the person moves closer than min_dist, the robot stops following.
max_linear

Maximum linear velocity of the robot.
max_angular

Maximum angular velocity of the robot.
max_frame_blocked

The robot will keep following the target for max_frame_blocked number of frames in the event of a temporary occlusion.
tracking_radius

The robot will keep following the target as long as the current target location = previous location +/- tracking_radius

Deployment Troubleshooting#

Failed to install Deb package: Please make sure to run sudo apt update before installing the necessary Deb packages.
You may stop the demo anytime by pressing ctrl-C.
If the robot rotates more than intended at each step, try reducing the parameter max_angular in the parameter file.
If the motor controller board does not start, restart the robot.
For general robot issues, refer to the troubleshooting guide.

ADBSCAN Follow-me#

Architecture & Algorithm#

ADBSCAN Algorithmic Foundation#

System Architecture#

Source Code & Workspace#

Intel®-Optimized ADBSCAN#

ADBSCAN Optimization Setup#

ADBSCAN Optimization Prerequisites#

Install and run the ROS 2 bag file Deb package#

Install and run optimized Deb package#

Install and run standard (unoptimized) Deb package#

Re-configurable parameters#

ADBSCAN Optimization Troubleshooting#

Simulate Follow-me in Gazebo#

Simulation Setup#

Simulation Prerequisites#

Install the Simulation Deb Package#

Activate Python Virtual Environment#

Install Python Modules#

Run Demo with 2D Lidar#

Run Demo with RealSense Camera#

Simulation Troubleshooting#

Deploy Follow-me on a Physical Mobile Robot (Clearpath Jackal / AAEON AMR)#

Deployment Setup#

Deployment Prerequisites#

Install the Deployment Packages#

Run Demo#

Deployment Troubleshooting#