Edge Video Classifier Retraining Using Confidence Thresholds
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing video processing systems face challenges in maintaining accurate object detection and classification due to changes in capture conditions over time, requiring significant computational resources and manual intervention for data retraining.
Innovation Solution
Implementing automated systems for periodic retraining of neural networks in edge video devices using confidence thresholds and local data analysis to identify false positives and negatives, reducing the need for human annotation and centralized processing.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If manual data review and centralized retraining are used to maintain model accuracy, then detection accuracy is improved, but computational resources and bandwidth requirements increase
Solution Approach 1:
The system enables edge devices to autonomously collect their own erroneous detection cases, annotate them locally using confidence thresholds, and perform self-retraining without requiring centralized manual review. This self-service approach maintains detection accuracy while eliminating the computational and bandwidth overhead of centralized processing
Solution Approach 2:
The retraining process is segmented into distributed micro-batches executed locally at each edge device rather than centralized bulk processing. Each device independently identifies and processes its own erroneous cases, dividing the overall retraining task into smaller, independent units that reduce computational resource concentration and bandwidth requirements
2Measurement precision
If frequent model updates are performed to adapt to environmental changes, then detection accuracy is improved, but the complexity of the retraining process increases
Solution Approach 1:
The system implements periodic retraining at edge devices based on accumulated erroneous cases rather than continuous or manual triggering. This periodic approach allows models to adapt to environmental changes over time while maintaining a simple, automated process that does not require complex scheduling or manual intervention
Solution Approach 2:
The system uses confidence thresholds to automatically identify and feedback erroneous detection cases to the retraining process. This feedback mechanism enables the system to self-correct accuracy decay without complex manual review processes, as low-confidence detections are automatically collected and used for targeted retraining
3Measurement precision
If centralized processing is used for model retraining, then model accuracy is improved, but network traffic increases
Solution Approach 1:
The system introduces local edge devices as intermediaries that perform data collection, annotation, and retraining locally before optionally syncing with centralized servers. This intermediary approach eliminates the need to transmit large volumes of raw video data to centralized processors, reducing network traffic while maintaining model accuracy through distributed processing
Data Source
AI summary
A digital video camera architecture for updating an object identification and tracking model deployed with the camera is disclosed. The invention comprises optics, a processor, a memory, and an artificial intelligence logic which may further comprise artificial neural networks. The architecture may identify objects according to the confidence threshold of a model. The confidence threshold may be monitored over time, and the model may be updated if the confidence threshold drops below an acceptable level. The data for retraining is ideally generated substantially internal to the camera. A classifier is generated to process the entire field data set stored on the camera to create a field data subset also stored on the camera. The field data subset may be run through the model to generate cases that may be used in further monitoring, training, and updating of the model. Classifiers may also be generated for images in different domains (e.g., lighting, weather, surveillance area, indoor, outdoor, urban, rural, etc.). These classifiers can be used to train the model to accurately identify objects and features independent of the domain of origin of the image being evaluated.