TL;DR
This paper presents a low-cost, infrastructure-free indoor positioning system for large spaces using smartphone cameras. It leverages deep learning and computer vision to detect common static objects (like doors) as landmarks, achieving sub-meter accuracy in a complex art museum environment.
What Problem Does This Paper Address?
The increasing demand for indoor location-based services (LBS) is hampered by the lack of a reliable, universally applicable solution, unlike outdoor GNSS. Existing indoor positioning methods often rely on expensive infrastructure, are limited in scale, or struggle with the complexity of large, visually rich environments like airports, malls, and museums. This paper specifically targets the challenge of accurate and cost-effective positioning within these large indoor spaces using readily available smartphone cameras.
Methods
The proposed system combines deep learning and computer vision techniques. Its core components are:
- Static Object Detection and Identification: They employ the Faster R-CNN algorithm, a well-established deep learning model for object detection. This network, built upon a VGG16 backbone, is trained to identify and localize common static objects (e.g., doors, windows) within smartphone images. The training process involves data augmentation to improve robustness.
- Control Point Coordinate Acquisition: Once static objects are identified, the system needs to establish reference points. This is done by defining "control points" on these static objects (e.g., corners of a door frame). In an offline phase, these control points' pixel coordinates are manually measured in "reference images." In the online phase, using SIFT feature extraction and RANSAC for outlier removal, the system matches features between the test image and reference images. A homographic matrix is then computed to map pixel coordinates from the reference image to the test image, thereby obtaining the pixel coordinates of the control points in the test image.
- Position Calculation:
- Position Estimation: A collinear equation model is used to relate the 3D space coordinates of control points to their 2D pixel coordinates in the image. With at least three control points, the smartphone's 3D position (X, Y, Z) can be iteratively estimated.
- Distance Estimation: To further refine accuracy and screen out gross errors, a distance estimation module is implemented. It calculates the distance between the smartphone and a static object using geometric principles and compares it to the estimated distance from the position calculation.
- Error Screening: A final check using a controlled error metric ensures the estimated position is acceptable.
Datasets
- Training Data: The paper mentions using "photos of static objects photographed from various angles at different distances" for training the Faster R-CNN network. They also utilized a dataset trained on ImageNet for transfer learning.
- Evaluation Data: Experiments were conducted in an art museum. The ground truth for control point coordinates was measured using a Hi-Target ZTS-420R Total Station with a reported 2 mm positioning error per 1000 m.
Key Insights and Ideas
- Leveraging Common Static Objects: The core idea of using ubiquitous, immovable objects like doors and windows as landmarks is a practical and cost-effective approach, avoiding the need for dedicated infrastructure.
- Imitating Human Spatial Cognition: The paper draws inspiration from how humans mentally map environments using visual cues, suggesting a natural and intuitive approach to indoor positioning.
- Deep Learning for Robust Detection: The use of Faster R-CNN for static object detection significantly improves robustness in complex visual environments, a common challenge for traditional computer vision methods.
- Web-Based System Architecture: The system is designed as web-based, with the smartphone acting as an image acquisition device and a server handling the computationally intensive algorithms. This minimizes smartphone battery drain and processing load.
- Low-Cost and Infrastructure-Free: The emphasis on using existing smartphone cameras and common objects makes the system highly accessible and scalable.
Key Results
- Positioning Accuracy: The system achieved positioning accuracy within 1 meter in a large art museum with a complicated visual environment.
- Distance Range: The system demonstrated effectiveness within a distance range of up to 40 meters.
- Static Object Recognition Performance: The Faster R-CNN model achieved high Accurate Precision (AP): 97.0% on average during training and 100% for testing across all identified static objects.
- Computational Cost: The static object recognition phase took approximately 0.3 seconds per image.
Limitations and Future Work
- Illumination and Shooting Angle Sensitivity: The performance can be influenced by variations in indoor lighting and shooting angles, which affect feature matching.
- Smartphone Distortion: Different smartphone cameras may introduce distortions that can impact position output, though camera calibration is suggested as a fix.
- Static Object Distribution: The distribution of control points on static objects is crucial. Improper distribution (e.g., along a straight line) can lead to higher positioning errors, as observed in two test points.
- Future Work: The authors plan to improve the system's robustness and experiment in other large indoor spaces with more complex layouts and topologies. They also suggest that non-transparent objects would be better for generating more feature points and increasing accuracy.
Summary updated