Multiple-Object Detection and Tracking via Identifier-Based Rounds
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Neural network-based object detection and tracking on edge compute devices face high latency and power consumption due to the need for continuous processing of all features for every object, exceeding the limited compute budget.
Innovation Solution
A method that assigns priorities and processes data elements sequentially based on unique identifiers, using a round robin algorithm to manage processing resources efficiently, selecting subsets of data elements for feature detection and tracking, and iteratively repeating this process until all elements are processed.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If continuous processing of all features for every object is performed, then tracking accuracy is improved, but power consumption and latency increase beyond the limited compute budget
Solution Approach 1:
The patent segments the continuous processing task into discrete processing rounds, where in each round only a subset of objects is processed. Objects are divided into groups based on their identifiers, and processing is distributed across multiple rounds rather than handling all objects simultaneously, thus reducing per-round power consumption while maintaining overall tracking accuracy.
Solution Approach 2:
The patent implements periodic processing where objects are processed in alternating rounds based on their identifiers. Even-indexed objects are processed in even rounds, odd-indexed objects in odd rounds. This periodic action distributes the computational load over time, reducing instantaneous power consumption while ensuring all objects are tracked over multiple cycles.
2Reliability
If all objects are processed simultaneously, then tracking completeness is improved, but processing latency increases due to limited compute budget
Solution Approach 1:
The patent segments the object set into multiple subsets using identifier-based grouping. Each processing round handles only one subset, dividing the total processing time into manageable segments. This segmentation reduces the time required per round while ensuring all objects are eventually processed across multiple rounds, improving overall tracking completeness without excessive latency.
Solution Approach 2:
The patent performs preliminary grouping of objects by their identifiers before processing begins. This preliminary action organizes objects into predetermined subsets, allowing the system to efficiently schedule and process different groups in alternating rounds without requiring complex real-time decisions, thus reducing processing latency while maintaining completeness.
3Measurement precision
If feature extraction is performed for every object in every frame, then detection accuracy is improved, but compute budget is exceeded
Solution Approach 1:
The patent applies local quality by differentiating processing intensity based on object identifiers. Instead of uniformly processing all objects with the same computational resources, the system selectively applies full feature extraction only to specific objects in specific rounds based on their identifier patterns, optimizing compute budget utilization while maintaining detection accuracy for tracked objects.
Solution Approach 2:
The patent implements partial action by performing feature extraction only on a subset of objects in each processing round rather than all objects. This partial processing approach reduces the computational load per round to fit within the compute budget, while the alternating round structure ensures that over time, all objects receive the necessary processing for accurate detection and tracking.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
According to an aspect of the present disclosure, there is provided a computer-implemented method comprising: receiving a stream of data containing a plurality of data elements, wherein each of the data elements is associated with a unique identifier; selecting a subset of said data elements using said identifiers; and processing the data elements of the selected subset to detect one or more features of the data elements of the selected subset.