Privacy-preserving visual tracking via head pose estimation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current visual tracking systems rely on facial recognition, which raises privacy concerns and are not effective in estimating gaze without face detection, especially in crowded environments where multiple cameras with different angles are needed to accurately track shopper behavior.
Innovation Solution
A system utilizing convolutional neural networks (CNNs) for real-time gaze and pose estimation without facial recognition, selecting the best camera based on head movements and field of view, and leveraging deep learning frameworks like Tensorflow for devices with limited resources, such as smartphones, to estimate yaw and pitch, ensuring privacy by not capturing face profiles.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If facial recognition is used for visual tracking, then tracking accuracy is improved, but privacy concerns increase and the system becomes less effective in crowded environments
Solution Approach 1:
The patent extracts and removes the facial recognition component from the visual tracking system, replacing it with head pose estimation based on landmarks. This eliminates the privacy concerns associated with facial recognition while maintaining tracking functionality through alternative biometric data (head position and orientation) that does not capture identifiable facial features.
Solution Approach 2:
The system segments the tracking task into two independent components: head pose estimation (using landmark detection) and gaze estimation (using eye position relative to head orientation). This segmentation allows accurate tracking without requiring full facial recognition, thus resolving the contradiction between tracking precision and privacy protection.
2Measurement precision
If multiple cameras with different angles are used to track shopper behavior in crowded environments, then tracking accuracy is improved, but device complexity increases
Solution Approach 1:
The patent introduces an intermediary processing layer that selects the most suitable camera view based on the detected landmark quality and viewing angle. Instead of processing data from all cameras simultaneously, the system acts as an intermediary to choose the optimal single view, thereby maintaining tracking accuracy in crowded environments while avoiding the complexity of processing multiple camera feeds in parallel.
Solution Approach 2:
The system dynamically adjusts camera selection based on real-time conditions such as shopper position, crowd density, and landmark detection quality. This dynamic adaptation allows the system to maintain high tracking accuracy by selecting the best camera angle for each specific situation without requiring a fixed complex multi-camera processing architecture.
3Measurement precision
If deep learning frameworks like Tensorflow are used for gaze and pose estimation, then estimation accuracy is improved, but computational resources required increase
Solution Approach 1:
The patent applies partial action by using a simplified version of deep learning models that process only the essential landmark coordinates for head pose and gaze estimation. Instead of using full facial recognition models that require extensive computational resources, the system uses streamlined neural networks that process only the necessary biometric data points, achieving sufficient accuracy while reducing energy consumption on mobile devices.
Data Source
AI summary
A visual tracker can be configured to obtain profile data associated with a pose of a living entity. In response to detecting a person, a camera can be selected from cameras. Additionally, in response to selecting the camera, the system can receive video data from the camera representative of a stance of the person. Consequently, the stance of the person can be estimated, resulting in an estimated stance.