Autonomous Parking Surveillance Rover

As part of the IFROS Lab at ELTE University, our team developed an autonomous parking surveillance application for Pomona, an AgileX Scout 2.0 rover. The system combines localization, navigation, perception, and mapping into a complete robotics stack deployed on a physical platform.

The rover builds an occupancy map of its environment, navigates through a parking area while avoiding dynamic obstacles, and detects vehicles and license plates with its onboard RGB-D camera. Detected license plates are localized in 3D and visualized alongside the environment map.

---
config:
    themeVariables:
        fontSize: "20px"
---
flowchart TD

    subgraph "`**Sensors**`"
        IN_CAMERA[/ RGB-D Camera /]
        IN_LIDAR[/ LIDAR /]
        IN_ODOMETRY[/ Wheel Odometry /]
        IN_GNSS[/ GNSS /] 
        IN_IMU[/ IMU /]
    end

    subgraph "`**Localization**`"
        ICP("KISS ICP")
        KALMAN("Kalman Filter Node")
    end

    subgraph "`**Navigation**`"
        GLOBAL_PLANNER("RRT*")
        CONTROLLER("Controller")
    end

    subgraph "`**Perception**`"
        DETECTOR("Car & License Plate Detector")
    end

    subgraph "`**Mapping**`"
        DETECTION_MAPPER("Detections to 3D Node")
        MAP("Octomap Server")
    end

    subgraph "`**Outputs**`"
        OUT_MAP[\ Map \]
        OUT_TWIST[\ Twist \]
    end


    IN_GNSS --> KALMAN 
    IN_ODOMETRY --> KALMAN
    IN_IMU --> KALMAN
    IN_CAMERA ---> DETECTOR
    IN_LIDAR --> MAP
    IN_LIDAR --> ICP

    ICP --> KALMAN
    GLOBAL_PLANNER --> CONTROLLER
    DETECTOR & KALMAN & IN_CAMERA --> DETECTION_MAPPER
    KALMAN --> MAP
    KALMAN --> GLOBAL_PLANNER

    MAP --> OUT_MAP
    DETECTION_MAPPER --> OUT_MAP
    CONTROLLER --> OUT_TWIST

Localization

An Extended Kalman Filter fuses wheel odometry, IMU data, GNSS, and LiDAR odometry from KISS-ICP. Combining these complementary sources provides a more stable pose estimate, particularly when wheel slip affects the mobile base.

The navigation stack uses OctoMap to represent obstacles and a custom Python implementation of RRT* to plan collision-free paths. A controller converts the planned trajectory into movement commands, while dynamic obstacle avoidance supports operation in changing environments.

The resulting occupancy map is augmented with estimated positions of detected vehicles and license plates, providing a spatial overview of the parking area.

Perception

Two YOLO-based detection models identify cars and license plates in the camera feed. Optical character recognition extracts the plate text, and the RGB-D data is combined with the robot pose to estimate the 3D position of each detection. An annotated camera stream is published for visualization.

Technology

Software: ROS, C++, Python, Extended Kalman Filter, KISS-ICP, YOLO, OCR, OctoMap
Hardware: AgileX Scout 2.0, NVIDIA Jetson Xavier, Velodyne LiDAR, Intel RealSense RGB-D camera, Xsens IMU and GNSS
Team: Samantha Caballero, Raul Musito, Lorenz Killer, Priyam Gupta

For the source code and project details, visit the SLAM Parking Inspector repository.

Autonomous Parking Surveillance Rover