360-Degree Image Edge Duplication for Neural Network Object Recognition
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current occupancy sensing methods in conference rooms, such as badge readers and Bluetooth Low Energy beacons, require attendees to take action or carry devices, leading to security issues and inefficient room usage tracking, as cameras may not be optimally positioned for security and occupancy analysis.
Innovation Solution
A 360-degree camera system combined with neural network object identification, which modifies 360-degree images by duplicating edge portions to create a modified rectangular format, allowing standard object recognition neural networks to accurately identify objects and track occupancy without re-training, enabling seamless conferencing experiences and enhanced security through facial recognition.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If standard neural networks are used for object recognition, then training efficiency is improved, but accuracy on 360-degree images deteriorates due to edge splitting
Solution Approach 1:
The patent segments the 360-degree image into multiple rectangular regions (e.g., left and right halves) and processes each region separately. This segmentation allows standard neural networks to accurately recognize objects without being confused by edge artifacts, while maintaining training efficiency using pre-existing models.
Solution Approach 2:
The patent introduces asymmetric edge handling by duplicating edge portions from one side to the other side of the image. This asymmetric modification creates symmetric representations of objects that span across the image boundary, enabling standard neural networks to accurately detect and recognize these objects without requiring retraining.
2Reliability
If cameras are positioned for optimal security monitoring, then security coverage is improved, but occupancy sensing capability deteriorates
Solution Approach 1:
The patent makes the camera system universal by enabling a single camera positioned for security monitoring to also perform occupancy sensing. The system processes 360-degree images from the security camera position to simultaneously detect both security-relevant objects and occupancy information, eliminating the need for separate cameras for each function.
Solution Approach 2:
The patent transitions from traditional 2D image analysis to 360-degree spherical image analysis. This dimensional change allows the camera to capture complete surround information from a single position, enabling both security monitoring and occupancy sensing without requiring multiple camera positions.
3Area of stationary object
If 360-degree images are processed directly, then complete room coverage is achieved, but object recognition accuracy deteriorates due to edge artifacts
Solution Approach 1:
The patent segments the 360-degree spherical image into multiple rectangular regions and processes each region independently. This segmentation maintains complete room coverage while eliminating edge artifacts that would otherwise reduce object recognition accuracy in the original 360-degree image format.
Solution Approach 2:
The patent copies edge portions from one side of the image to the other side to create duplicated edge representations. This copying technique preserves complete room coverage information while creating artifact-free edges that improve object recognition accuracy for standard neural networks.
Data Source
AI summary
A method, according to one example, includes receiving a 360-degree image that was captured by a 360-degree camera, converting the 360-degree image into a rectangular image, and copying an edge portion from a first edge of the rectangular image and pasting the copied edge portion to a second edge of the rectangular image, thereby forming a modified rectangular image. The method further includes applying a neural network to the modified rectangular image to identify objects appearing in the modified rectangular image, wherein the modified rectangular image facilitates object identification near the second edge.