Automotive Global Motion Modeling for Robust Optical Flow
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional global motion models are ineffective in handling video captured from a moving vehicle, especially due to high-speed motion, wide-angle lenses, and challenging environmental conditions like low light and limited texture.
Innovation Solution
A method that uses a motion model developed for automotive applications, incorporating parameters such as camera orientation, lens distortion, and inertial measurement units to derive a dense global motion field, which can be used to augment optical flow calculation through fusion, search candidate guidance, pre-warping, and modified cost functions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If conventional global motion models are used for automotive video, then the models can handle general video processing, but they fail to effectively handle high-speed motion and wide-angle lens distortions specific to automotive applications
Solution Approach 1:
The patent applies local quality by creating a specialized global motion model tailored specifically for automotive conditions rather than using a general-purpose model. The model incorporates automotive-specific parameters such as vehicle speed, steering angle, and suspension motion to accurately represent the unique motion patterns encountered in automotive video, thereby achieving both adaptability to specific conditions and reliability in motion estimation.
Solution Approach 2:
The patent changes the parameters of the global motion model to accommodate automotive-specific conditions. It introduces parameters such as vehicle velocity, steering wheel angle, and suspension displacement that are specific to automotive applications. These parameter changes enable the model to effectively handle high-speed motion and wide-angle lens distortions that conventional models cannot address.
2Device complexity
If image-based motion analysis is used in challenging environmental conditions, then the analysis can be performed without additional sensors, but the estimation becomes unreliable due to low light, weather conditions, shadows, and limited texture
Solution Approach 1:
The patent applies preliminary action by pre-characterizing the camera system with intrinsic parameters (focal length, principal point, distortion coefficients) and establishing the relationship between vehicle motion and image plane motion before actual motion analysis is performed. This pre-characterization enables the system to reliably estimate motion parameters even in challenging environmental conditions, without requiring additional sensors or increasing system complexity.
3Adaptability or versatility
If multiple camera orientations and lens distortions are considered in machine learning approaches, then the models can handle diverse camera configurations, but significant complexity arises requiring different models for each configuration
Solution Approach 1:
The patent achieves universality by creating a single, unified global motion model that can handle multiple camera orientations and lens distortion types. The model uses general parameters such as camera intrinsic parameters (which can be determined through calibration) and vehicle motion parameters to adapt to different camera configurations without requiring separate specialized models, thereby maintaining low complexity while achieving high versatility.
4Measurement precision
If high-accuracy high-resolution motion information is recorded from real outdoor scenes, then training data quality improves, but the methods for recording such data are limited and difficult to obtain
Solution Approach 1:
The patent applies self-service by using the vehicle's own motion sensors (GPS, inertial measurement units, steering angle sensors) to generate ground truth motion data during normal operation. This eliminates the need for specialized data collection equipment or controlled environments, enabling the system to accumulate high-precision training data easily from real-world driving conditions without requiring additional manufacturing complexity.
Data Source
AI summary
Certain aspects of the present disclosure provide techniques for global motion modeling. Embodiments include receiving a first image and a second image from a camera attached to a moving object. Embodiments include identifying a pixel in the first image. Embodiments include determining, based on one or more parameters associated with the camera, a vector representing a range of locations in which a real-world point corresponding to the pixel is likely to be found in the second image, wherein the parameters associated with the camera comprise: a first parameter related to a location of the camera relative to a ground surface; a second parameter related to motion of the moving object; and a third parameter related to an orientation of the camera relative to the ground surface. Embodiments include determining, using the vector, a location of the real-world point in the second image.