Robot Device Mode Switching for Privacy-Aware Object Recognition
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing robot devices face challenges in balancing privacy protection and efficient object recognition, as they must either transmit sensitive images to cloud servers for high-performance machine learning, risking privacy, or rely on limited on-device processing, which may lead to inaccurate object recognition and operational issues.
Innovation Solution
A robot device that switches between a cloud machine learning model and an on-device machine learning model based on the presence of a person in its driving area, using the cloud model for privacy protection when no person is detected and the on-device model for faster processing when a person is present, thereby ensuring privacy and improving recognition accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If cloud machine learning model is used for object recognition, then recognition accuracy is improved, but user privacy is compromised due to image transmission
Solution Approach 1:
The system dynamically switches between cloud-based and on-device machine learning models based on real-time detection of persons in the driving area. When a person is detected, the system transitions to on-device processing to protect privacy; when no person is detected, it uses cloud processing for higher accuracy. This dynamic adaptation resolves the contradiction between accuracy and privacy.
Solution Approach 2:
The system applies different processing qualities to different spatial contexts: cloud-based high-accuracy processing for areas without persons, and on-device privacy-preserving processing for areas with persons. This local differentiation allows the system to optimize accuracy where safe and protect privacy where needed.
2Object-affected harmful factors
If on-device machine learning model is used for object recognition, then privacy is protected, but recognition accuracy deteriorates
Solution Approach 1:
The system dynamically selects the appropriate machine learning model based on environmental conditions. On-device models are activated only when privacy is needed (person detected), while cloud models are used when maximum accuracy is required (no person detected). This dynamic selection ensures both privacy protection and high accuracy are achieved at different times.
Solution Approach 2:
A person detection mechanism acts as an intermediary that determines which processing path to take. This intermediary evaluates the privacy risk and routes images to either on-device or cloud processing accordingly, ensuring that privacy-protective measures are applied only when necessary.
3Power
If cloud server is used for processing, then processing power is sufficient, but data transmission time and network dependency increase
Solution Approach 1:
The system segments processing tasks into two categories: time-critical processing done on-device and computation-intensive processing done on cloud. By segmenting based on urgency and resource requirements, the system minimizes transmission time for urgent tasks while still leveraging cloud power for non-urgent heavy processing.
Solution Approach 2:
The system performs preliminary person detection and mode selection before image transmission. By determining in advance whether cloud processing is needed, the system avoids unnecessary data transmission and its associated time delays, only transmitting images when cloud processing is actually required.
Data Source
AI summary
A robot device includes at least one processor configured to detect a person in a driving area of the robot device, based on a determination that no person is present in the driving area, recognize an object in an input image generated from the image signal using a cloud machine learning model, in a first mode, based on a determination that a person is present in the driving area, recognize the object in the input image generated from the image signal using an on-device machine learning model, in a second mode, and control the driving of the robot device through the moving assembly by using a result of recognizing the object, wherein the cloud machine learning model operates on a cloud server connected through the communication interface, and the on-device machine learning model operates on the robot device.