Multi-Camera Object Detection via Client-Side Feature Extraction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing object detection technologies face challenges in aggregating data from multiple sources to train general object detection models while ensuring the privacy of source data, particularly in multi-camera surveillance systems.
Innovation Solution
A multi-camera domain adaptive object detection system and method that involves a server-end device and multiple client-end devices, where the client-end devices train a global model using both target and source data, extract feature data, and transmit only feature data to the server-end device for assembling loss functions and retraining the global model.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If source data from multiple cameras is transmitted to a centralized server for training, then the model can be trained with diverse data, but data privacy and security are compromised
Solution Approach 1:
The patent extracts only the necessary feature representations from the source data at client devices, rather than transmitting the complete source data to the server. This extraction process removes privacy-sensitive information while retaining the essential characteristics needed for model training, thereby resolving the contradiction between data diversity and privacy protection.
Solution Approach 2:
The patent introduces feature representations as an intermediary between the source data and the training process. Instead of directly transmitting source data, the system converts it into feature representations that serve as a mediator, enabling model training with diverse data while preventing direct access to privacy-sensitive source information.
2Quantity of substance
If large amounts of image data are transmitted to a centralized server, then comprehensive training data is available, but transmission volume and network bandwidth requirements increase significantly
Solution Approach 1:
The patent extracts only the essential feature representations from the source image data at the client devices before transmission. This extraction eliminates redundant and unnecessary data components, significantly reducing the transmission volume while preserving the critical information needed for effective model training.
Solution Approach 2:
The patent transmits only the necessary portion of the source data (feature representations) rather than the complete dataset. This partial transmission approach provides sufficient information for training purposes while avoiding the excessive bandwidth consumption that would result from transmitting full-resolution images and metadata.
3Object-affected harmful factors
If feature data is extracted and transmitted instead of source data, then data privacy is protected and transmission volume is reduced, but the complexity of the training process increases
Solution Approach 1:
The patent performs feature extraction as a preliminary action at the client devices before data transmission. By preprocessing the source data into feature representations in advance, the system simplifies the subsequent training process at the server, as the features are already extracted and ready for use, thereby reducing the overall computational complexity despite the added extraction step.
Data Source
AI summary
A multi-camera domain adaptive object detection system and detection method are provided. A server-end device of the system includes a target camera for obtaining target data. Each client-end device of the system includes a source camera for obtaining source data. The system executes procedures of: the server-end device transmits a global model to each client-end device; each client-end device trains the global model according to the target data and corresponding source data to obtain a trained model; each client-end device inputs the target data into the trained model, extracts feature values to obtain a feature data, and transmits each feature data to the server-end device; the server-end device assembles the feature data to obtain multiple loss functions; and the server-end device trains the global model according to the target data and loss functions to obtain a trained global model.