Asynchronous Object ROI Detection in Video Frames
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for detecting regions of interest (ROIs) in images, such as surveillance systems, require long processing times, high power consumption, and large output transfer bandwidths due to the need for simultaneous inference using multiple object detectors on a single frame.
Innovation Solution
The implementation of an asynchronous inference method that determines partial ROIs from each frame in a video, switching control between different object detectors and processing regions, reducing computational costs to approximately 1/(N*M) of simultaneous inference methods, where N is the number of categories and M is the number of processing regions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If simultaneous inference using multiple object detectors is applied to a single frame, then all required ROIs can be detected from one frame, but processing time increases, power consumption increases, and output transfer bandwidth increases
Solution Approach 1:
The patent divides the image into multiple processing regions and assigns different object detectors to different regions asynchronously. Instead of applying all detectors to the entire image simultaneously, each detector processes only its assigned region, thereby reducing the computational overlap and processing time while maintaining comprehensive ROI detection coverage.
Solution Approach 2:
The system dynamically switches between different object detectors based on the current frame's characteristics and control rules. Rather than statically applying all detectors to every frame, the system adaptively selects which detector to apply to which region in each frame, optimizing processing speed while maintaining detection accuracy.
2Measurement precision
If simultaneous inference using multiple object detectors is applied to a single frame, then all required ROIs can be detected from one frame, but power consumption increases
Solution Approach 1:
The patent segments the computational workload by dividing the image into multiple regions and assigning different detectors to different regions. This segmentation reduces the total computational load per frame by eliminating redundant processing across overlapping regions, thereby reducing power consumption while maintaining complete ROI detection.
Solution Approach 2:
The system applies object detectors selectively to partial regions of the image rather than processing the entire image with all detectors. This partial action approach reduces the total number of detection operations performed per frame, lowering power consumption while still achieving comprehensive ROI detection through the coordinated work of multiple detectors on different regions.
3Measurement precision
If simultaneous inference using multiple object detectors is applied to a single frame, then all required ROIs can be detected from one frame, but output transfer bandwidth increases
Solution Approach 1:
The patent segments the detection output by region and detector, allowing results to be processed and transmitted in a more organized manner. By dividing the computational workload and corresponding output across multiple regions rather than generating all results simultaneously from a single full-image processing pass, the system reduces peak bandwidth requirements.
4Productivity
If asynchronous inference is applied to video frames, then processing time, power consumption, and transfer bandwidth are reduced, but the method requires coordination between multiple detectors and frames
Solution Approach 1:
The patent establishes control rules in advance that define which object detector should be applied to which processing region for each frame in the video sequence. By pre-configuring the detection strategy rather than making real-time decisions about detector selection, the system reduces runtime coordination complexity while maintaining the efficiency benefits of asynchronous processing.
Data Source
AI summary
An image data processing method includes receiving, from an image sensor, frame image data of a plurality of frames; receiving a plurality of control rules comprising a respective control rule for each of the frames, wherein each of the control rules identifies one of a plurality of process regions and one of a plurality of object detectors; identifying a region of interest in each frame by a location and a category, comprising applying the object detector identified by the respective control rule to the respective frame image data in the process region identified by the respective control rule; identifying a final region of interest based on the identified regions of interest; and reporting the final region of interest.