Clustering-Based Index for Real-Time Video Face Lookups
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional indexing methods for real-time video surveillance face challenges in maintaining sub-second look-up times and controlling index size growth as the number of faces increases, leading to high memory and CPU requirements or offline processing, which affects performance and accuracy.
Innovation Solution
A clustering-based index structure using k-medoids clustering assigns faces to unique clusters with centroids, allowing for sub-linear/logarithmic index size growth and enabling sub-second look-ups by comparing query feature vectors with cluster centroids and members for similarity above a threshold.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Speed
If conventional indexing methods are used, then look-up speed is maintained, but index size grows linearly with the number of faces
Solution Approach 1:
The patent segments the face collection into multiple clusters, each with its own centroid. Instead of maintaining a single large index of all faces, the system divides the data into smaller groups (clusters) that can be independently managed and searched. This segmentation allows the index size to grow sub-linearly while maintaining efficient look-up performance through hierarchical searching.
Solution Approach 2:
The patent introduces cluster centroids as intermediary elements between the query and the actual face data. Rather than directly comparing queries against all individual faces, the system first compares queries against cluster centroids (which are compact representations), and only then proceeds to compare against individual faces within relevant clusters. This intermediary approach dramatically reduces the search space and index requirements.
2Quantity of substance
If index size is reduced to control memory usage, then memory requirements decrease, but look-up accuracy deteriorates
Solution Approach 1:
The patent creates compact copies of face data in the form of cluster centroids, which serve as simplified representations of entire clusters. These centroids capture the essential characteristics of multiple faces in a condensed form, allowing the system to maintain accurate look-up capability while using minimal memory. The centroids act as proxies that preserve the most important information needed for accurate matching.
Solution Approach 2:
The patent changes the representation parameters from storing complete face data to storing aggregated cluster characteristics (centroids). This parameter transformation allows the system to work with much smaller data structures while maintaining the ability to perform accurate comparisons. The centroids encode sufficient information to guide the search process and identify relevant clusters for detailed comparison.
3Productivity
If more CPU resources are allocated, then real-time processing capability improves, but operational cost increases
Solution Approach 1:
The patent segments the processing workload by dividing the face collection into clusters. This segmentation allows the system to process queries more efficiently by only examining relevant clusters rather than all faces, reducing the computational burden and enabling real-time processing with fewer CPU resources.
Solution Approach 2:
The patent performs preliminary clustering and centroid computation before the actual look-up process. This preliminary organization of data into clusters with pre-computed centroids enables faster query processing, as the system can quickly identify relevant clusters without performing exhaustive comparisons, thereby reducing real-time computational requirements and operational costs.
Data Source
AI summary
Systems and methods for recognizing a face are disclosed and includes receiving images of faces; generating feature vectors of the images; generating clusters of feature vectors each with a centroids or a cluster representative; for a query to search for a face, generating corresponding feature vectors for the face and comparing the feature vector with the centroids of all clusters; for clusters above a similarity threshold, comparing cluster members with the corresponding feature vector; and indicating as matching candidates for cluster members with similarity above a threshold.