QuadricSLAM: Dual Quadrics from Object Detections as Landmarks in Object-oriented SLAM (1804.04011v3)

Abstract: In this paper, we use 2D object detections from multiple views to simultaneously estimate a 3D quadric surface for each object and localize the camera position. We derive a SLAM formulation that uses dual quadrics as 3D landmark representations, exploiting their ability to compactly represent the size, position and orientation of an object, and show how 2D object detections can directly constrain the quadric parameters via a novel geometric error formulation. We develop a sensor model for object detectors that addresses the challenge of partially visible objects, and demonstrate how to jointly estimate the camera pose and constrained dual quadric parameters in factor graph based SLAM with a general perspective camera.

• The paper introduces dual quadrics as a novel landmark parameterization to efficiently capture object size, position, and orientation in 3D space.
• It integrates state-of-the-art object detectors with a geometric error formulation to accurately derive landmark constraints from 2D bounding boxes.
• The method employs a factor graph-based SLAM formulation that jointly estimates camera poses and object parameters even in cluttered, partially occluded environments.

Overview of "QuadricSLAM: Dual Quadrics from Object Detections as Landmarks in Object-oriented SLAM"

The paper "QuadricSLAM: Dual Quadrics from Object Detections as Landmarks in Object-oriented SLAM" presents a novel approach to enhance Simultaneous Localization and Mapping (SLAM) by introducing semantically meaningful, object-oriented 3D maps through the use of dual quadrics as landmark representations. This work is motivated by recent advancements in vision-based object detection utilizing Convolutional Neural Networks (ConvNets) and addresses existing gaps in SLAM’s capability to incorporate semantic scene understanding.

Key Contributions

The research makes several important contributions to the SLAM literature:

1. Dual Quadrics as Landmark Parameterization: The paper introduces the concept of using dual quadrics for object representation in SLAM. Quadrics provide a compact and efficient way to represent an object's size, position, and orientation in 3D space, making them a robust choice for semantically enriching SLAM systems without relying on pre-existing CAD models of objects.
2. Integration of Object Detectors: The authors demonstrate the integration of modern object detection systems, such as YOLOv3, as sensors for SLAM. They propose a novel geometric error formulation that constrains dual quadric parameters directly from 2D object detection bounding boxes, a crucial step in enabling SLAM systems to leverage the bounding box data for accurate object localization and mapping.
3. Factor Graph-Based SLAM Formulation: A factor graph-based SLAM formulation is developed, which jointly estimates camera poses and dual quadric parameters. This approach is robust to partially visible objects and employs a general perspective camera model, thereby enhancing the applicability of SLAM systems in realistic environments, including indoor and cluttered scenarios.
4. Geometric Error Formulation: The research evaluates the traditional algebraic error formulations for quadric projection against their novel geometric error term, finding the latter more robust to scenarios with occluded or partially visible objects. This advancement improves the reliability of quadric parameter estimation under typical conditions faced in robotic vision applications.

Experimental Validation and Results

The authors conduct extensive evaluations in both real-world and simulated environments:

• TUM RGB-D Dataset: Real-world experiments on challenging sequences from the TUM RGB-D dataset revealed that the approach improves the trajectory estimation over standard visual odometry techniques. While slightly falling behind the state-of-the-art ORB-SLAM2 in some scenarios, QuadricSLAM demonstrates a significant advancement in developing semantically meaningful maps by integrating object-level semantics.
• High-Fidelity Simulation: In a controlled simulation environment, QuadricSLAM displayed substantial improvements over noisy odometry data in both trajectory accuracy and landmark estimation. The results underscore the effectiveness of object-oriented landmarks in correcting significant localization errors.

Implications and Future Directions

This research provides a significant step toward enriching SLAM maps with object-level semantics, thereby enhancing the utility of robotic systems in scenarios that demand greater scene understanding and interaction complexity. The work facilitates a more intuitive integration between detected objects and SLAM, paving the way for robust applications in autonomous navigation, surveillance, and augmented reality environments.

Future research could explore expanding the method with richer object detection confidence measures, improved handling of occlusions, and considering the integration of additional sensory data like depth to reject spurious detections. As SLAM systems increasingly incorporate semantic understanding, leveraging dual quadrics may further enable robots to draw meaningful inferences and act more intelligently in dynamic environments.