Target Object Tracking With Frame Skipping and Motion Estimation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing tracking solutions based on deep learning models are energy-intensive and computational resource-intensive, limiting their deployment in real-world applications, particularly in video surveillance systems with numerous cameras.
Innovation Solution
A method for tracking a target object by processing only a portion of the image stream at a detection frequency lower than the capture frequency, using a tracking phase that skips processing intermediate images, combined with position estimation based on linear movement between processed images.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If all images in the image stream are processed by deep learning models for tracking, then tracking accuracy is improved, but energy consumption and computational resource usage increase significantly
Solution Approach 1:
The patent applies partial action by processing only a subset of images rather than all images in the stream. It uses a decimation factor to select every Nth image for processing, combined with motion detection to identify which specific images within that subset require tracking analysis. This partial processing approach maintains reliable tracking by focusing computational resources on images containing moving objects while skipping static scenes, thereby significantly reducing energy consumption.
Solution Approach 2:
The patent segments the image processing task into multiple stages: first applying a decimation filter to select candidate images, then applying motion detection to further filter candidates, and finally applying deep learning tracking only to images containing motion. This segmentation allows the system to maintain tracking accuracy for relevant frames while avoiding unnecessary processing of static images, thus resolving the contradiction between accuracy and energy consumption.
2Reliability
If all images in the image stream are processed by deep learning models for tracking, then tracking accuracy is improved, but computational resource usage increases significantly
Solution Approach 1:
The patent reduces computational resource usage by applying partial action - processing only a fraction of images through the computationally intensive deep learning tracking model. By using a decimation factor to select every Nth image and further filtering with motion detection, the system maintains tracking accuracy for dynamic scenes while avoiding unnecessary computation on static frames, thereby reducing overall computational resource requirements.
Solution Approach 2:
The patent segments the computational workload into hierarchical stages: a coarse decimation filter that reduces image frequency, a medium-level motion detection filter that identifies dynamic content, and a fine-level deep learning tracking application. This segmentation strategy maintains tracking accuracy by ensuring deep learning is applied only when necessary, while significantly reducing the total computational burden compared to processing every image through the full pipeline.
3Use of energy by moving object
If the detection frequency is reduced to lower than capture frequency, then energy consumption and computational resource usage decrease, but tracking precision may be affected
Solution Approach 1:
The patent applies partial action by selectively processing only those images that contain motion, rather than uniformly processing all images at capture frequency. The motion detection mechanism identifies frames with moving objects and triggers tracking processing only for those frames, even though the overall detection frequency is lower than capture frequency. This selective approach maintains tracking precision for dynamic targets while reducing energy consumption by skipping static frames.
Solution Approach 2:
The patent applies preliminary action by performing motion detection as a pre-filter before applying the computationally intensive tracking algorithm. This preliminary check identifies which images warrant further processing, allowing the system to maintain tracking precision for moving objects while avoiding unnecessary computation on static scenes, thereby reducing energy consumption without sacrificing precision for relevant targets.
Data Source
AI summary
The invention relates to a method of tracking a target object in an image stream captured by a camera, at a capture frequency. The method includes a tracking phase, implemented individually for several processed images of said image stream, wherein the tracking phase includes detecting at least one object, and its position, in the processed image, and identifying the target object among the at least one object detected in the processed image. The tracking phase is carried out at a detection frequency, that is lower than the capture frequency, so that two images processed during two successive iterations of the tracking phase are separated by at least one non-processed image to which the tracking phase is not applied. The invention also relates to a computer program, a device, and a system implementing such a method.