Wide-Angle Image Processing With Virtual Views for Low-Loss Vision
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional wide-angle cameras, such as fisheye cameras, distort images and require extensive image resizing, leading to information loss and making it difficult for computer vision algorithms to recognize objects, while traditional perspective cameras have limited field of view and necessitate multiple installations for large areas.
Innovation Solution
A system and method that processes wide-angle images by generating multiple virtual images with minimal information loss, using defined target buffers and virtual camera views to match input parameters of computer vision models, allowing for efficient object detection and classification.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Area of stationary object
If wide-angle cameras are used to cover large areas, then the field of view is improved, but image distortion and warping increase
Solution Approach 1:
The patent divides the single wide-angle fisheye image into multiple virtual camera views (typically 4-9 views), each representing a specific field of view direction. This segmentation transforms one highly distorted image into multiple less-distorted virtual images that can be processed separately by computer vision algorithms, resolving the contradiction between wide coverage and image quality.
2Adaptability or versatility
If fisheye images are scaled down to match model input size, then compatibility with computer vision models is improved, but information loss increases
Solution Approach 1:
Instead of scaling down the entire high-resolution fisheye image, the patent segments it into multiple virtual camera views. Each virtual view is then scaled down individually to match model input dimensions (e.g., 416x416 for YoloV3). This approach preserves more information because each virtual view represents a focused region rather than compressing the entire wide-angle scene into a small size.
Solution Approach 2:
The patent transforms the spatial dimension problem by creating multiple virtual views from different angular perspectives. This dimensional transformation allows the system to maintain higher resolution in each view while still achieving model compatibility, effectively trading one dimension (angular coverage) for another (resolution preservation).
3Manufacturing precision
If de-warping or rectification is performed on the entire fisheye image, then image distortion is reduced, but the resulting image still has strong curvature and is not suitable for computer vision algorithms
Solution Approach 1:
The patent applies segmentation by dividing the fisheye image into multiple virtual camera views, each with a limited field of view. This segmentation inherently reduces distortion in each view compared to the entire fisheye image, and the resulting virtual views are naturally more suitable for standard computer vision algorithms without requiring aggressive de-warping that would preserve unwanted curvature.
4Area of stationary object
If multiple cameras are installed to cover large areas, then coverage area is improved, but system complexity increases
Solution Approach 1:
The patent makes a single fisheye camera perform the function of multiple cameras by generating multiple virtual camera views from one physical sensor. Each virtual view can be processed independently by computer vision models, achieving multi-camera coverage functionality with single-camera hardware, thus reducing system complexity while maintaining wide area coverage.
Data Source
AI summary
Disclosed is a system (100) for processing images related to a target region. The system comprises a memory (102) and a computing arrangement (104) in communication with the memory. The computing arrangement is configured to define one or more target buffers (106) in the memory of a size corresponding to one or more image parameters to be used as an input to a model implemented for the processing the images. The computing arrangement is further configured to receive at least one image (300, 400A, 400B, 400C, 400D) of the target region from a wide-angle camera (108, 202), define two or more virtual camera views (302, 304, 306, 412, 416, 418, 420, 422) and process the at least one image to generate two or more virtual images (308, 310, 312, 400E) based on the defined one or more target buffers and the defined two or more virtual camera views. The computing arrangement is further configured to implement the said model to analyse the generated two or more virtual images.