Object Tracking Cost Matrix for Embedded Video Analytics
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing video analytics solutions for object tracking in images and videos face high computational complexity, making them unsuitable for embedded platforms, despite their need for accurate and efficient tracking.
Innovation Solution
An object tracking system that computes a cost function based on visual dissimilarity, physical distance, and expected location between objects in one frame and candidate objects in another frame, using a processor and memory unit to store and select pairs of objects for accurate tracking.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If online learning based strategies with discriminative classifier models (such as SVM) are used for object tracking, then tracking accuracy is improved, but computational complexity increases making it unsuitable for embedded platforms
Solution Approach 1:
The patent segments the tracking problem into distinct components: object detection in current frame, candidate object detection in next frame, cost function computation (with sub-components for visual dissimilarity, physical distance, and expected location), and optimal pairing selection. This segmentation allows each component to be optimized independently for embedded platforms while maintaining overall tracking accuracy.
Solution Approach 2:
The patent changes the parameters used in tracking from complex discriminative classifier models to a cost function based on visual dissimilarity, physical distance, and expected location. This parameter transformation reduces computational complexity while preserving tracking accuracy by focusing on physically meaningful metrics that can be computed efficiently on embedded devices.
2Reliability
If complex discriminative classifier models are continuously learned and updated, then object separation from surroundings is improved, but computational resources required increase
Solution Approach 1:
The patent replaces expensive, continuously learned discriminative classifier models with a simpler cost function approach that can be computed efficiently without continuous learning. The cost function uses readily available visual features and physical metrics that can be evaluated cheaply and repeatedly, eliminating the need for resource-intensive model training and updates while maintaining reliable object separation.
3Ease of manufacture
If MeanShift or Optical Flow techniques are used for object tracking, then implementation simplicity is improved, but tracking precision deteriorates
Solution Approach 1:
The patent merges multiple tracking considerations into a unified cost function that combines visual dissimilarity (from MeanShift-like approaches), physical distance (from Optical Flow-like approaches), and expected location predictions. This merging preserves the simplicity of individual techniques while achieving superior tracking precision through their coordinated integration in the Hungarian algorithm framework.
Data Source
AI summary
The present invention discloses methods, systems, and computer programmable products for tracking objects across a first frame and a second frame of a video. An object tracking system computes a cost function between each object in the first frame and each object in the second frame. Further, a pair of objects is selected from the first frame and the second frame based on a pre-defined criteria which is based on the computed cost function. Thereafter, the object tracking system established a correspondence between the selected pair of objects.