Abstract
Analyzed the quality and its influence factors of relative orientation of vehicle-based sequence images by comparing with the result obtained from Position & Orientation System using Direct Geo-referencing. Studies have shown that, under normal circumstances, the image Relative Orientation based on Computer vision is more robust, and under special conditions, it is more practical for analyzing and increasing the quality of vision measurement. Besides these, the method can be borrowed by many surveying and mapping related fields, taking indoor robot mobile environment awareness and unmanned automobile for instance.
Introduction
Mobile Mapping Systems (MMS) rely on knowing the exact position and orientation of a vehicle to create accurate maps. Traditionally, this is achieved using expensive hardware like Inertial Measurement Units (IMUs) and GPS. While photogrammetry offers a software-based alternative called Relative Orientation (RO), traditional models often fail when a vehicle makes significant turns (large heading changes). These older models rely on iterative "Least Squares" methods that can get stuck in local optima if the initial guess isn't perfect.
This paper evaluates the accuracy of Computer Vision (CV) techniques for determining the relative orientation of vehicle-based sequence images. The authors set out to prove that CV-based RO—specifically using matrix decomposition—is a more robust and practical alternative for modern mobile mapping, indoor robotics, and autonomous driving.
Methods
The authors propose a pipeline that moves away from traditional photogrammetric resection toward a modern CV approach. The process involves several key stages:
- Feature Extraction & Matching: Instead of gray-correlation matching, which struggles with rotation, the SIFT (Scale-Invariant Feature Transform) algorithm is used for its robustness against changes in scale, rotation, and illumination.
- Outlier Rejection: RANSAC (Random Sample Consensus) filters out false matches, ensuring only high-quality point pairs are used for subsequent calculations.
- Matrix Estimation: The Fundamental Matrix ($F$) is calculated using an 8-point algorithm, followed by the derivation of the Essential Matrix ($E$) using camera calibration parameters.
- SVD Decomposition: The rotation matrix ($R$) is extracted from the Essential Matrix via Singular Value Decomposition (SVD), providing Euler angles (yaw, pitch, and roll) without complex iterative linearization.
- Validation: Results are compared against data from a high-end Applanix POS/LV system.
Datasets and Experimental Setup
The study utilizes a real-world dataset provided by the Beijing Maggroup Group. The hardware setup included a camera with 1600x1200 resolution and a 7.5mm focal length. The test environment consisted of 11 sequential images captured around an urban supermarket, featuring heading angles ranging from 3° to 17°, which represent typical urban driving maneuvers.
Key Insights and Analysis
- The SIFT Bottleneck: While SIFT is highly stable, it is computationally expensive. Feature matching accounted for approximately 90% of the total processing time, taking about 10 seconds per image pair.
- Systematic Misalignment: A consistent difference between CV-calculated and IMU-measured angles suggests a systematic error caused by the physical misalignment between the camera and IMU coordinate systems, rather than algorithmic failure.
- Geometric Stability: Interestingly, larger heading changes resulted in smaller compute errors. Conversely, very small heading changes led to instability, suggesting that CV methods thrive when there is significant geometric "movement" between frames.
Key Results
The CV-based relative orientation demonstrated high reliability, achieving an absolute error within 3°. The Root Mean Square (RMS) error for the heading angle was 0.768°, and RANSAC filtering kept the incorrect matching rate within 2%.
As shown in the comparison graph below, the CV-calculated heading values followed the exact same trajectory or "tendency" as the high-precision IMU data, validating the method's reliability for autonomous navigation.
Limitations and Future Work
Despite the promising results, several challenges remain:
- Processing Speed: The current 10-second processing time per pair is unsuitable for real-time applications, necessitating future hardware acceleration.
- Calibration Dependence: High-quality RO is heavily dependent on rigorous camera calibration to minimize systematic errors.
- Overlap Requirements: The method requires high image overlap (ideally >60%), which may limit vehicle speed or require higher frame-rate cameras.
- Environmental Factors: Occlusion and lighting changes continue to impact the precision of feature extraction.
Summary updated