Moving Object Imaging with CNN-Guided Camera AX Control
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing computer vision techniques struggle to capture high-quality images of moving objects without requiring extensive training data or hand-crafted features, particularly in dynamic environments like satellite separation in space or military drone operations, where objects of interest are not known beforehand.
Innovation Solution
A method using pre-trained convolutional neural networks for object tracking, generating candidate bounding boxes, and calculating feature vectors to dynamically control auto-exposure, auto-focus, and auto-white balance algorithms based on object movement, without the need for prior training data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional computer vision techniques (background subtraction, optical flow, template matching) are used for object tracking, then object location can be determined in image frames, but the ability to capture high-quality images of moving objects is limited without extensive training data or hand-crafted features
Solution Approach 1:
The system performs preliminary object detection and bounding box generation before image capture, using pre-trained convolutional neural networks to identify objects of interest in advance. This allows the system to prepare tracking parameters and capture settings beforehand, ensuring both accurate object location and consistent image quality without requiring extensive runtime training data
Solution Approach 2:
The patent replaces conventional mechanical computer vision techniques (background subtraction, optical flow, template matching) with a neural network-based system. The pre-trained convolutional neural networks automatically learn features from data, eliminating the need for hand-crafted features while providing both accurate tracking and reliable image quality through learned representations
2Productivity
If automated algorithms (AX algorithms) are controlled dynamically based on object statistics, then real-time image optimization for viewing the object is enabled, but the system complexity increases
Solution Approach 1:
The pre-trained convolutional neural networks serve multiple functions simultaneously: they perform object detection, generate bounding boxes, extract object statistics, and provide features for tracking. This multi-functionality enables real-time image optimization through AX algorithms while avoiding the need for separate dedicated components for each function, thereby managing system complexity
Solution Approach 2:
The neural network system automatically generates object statistics and bounding box information that directly controls the AX algorithms. The system serves itself by using the same neural network infrastructure for both object identification and capture parameter control, eliminating the need for separate control systems and reducing overall complexity
3Ease of manufacture
If pre-trained convolutional neural networks are used for object tracking and bounding box generation, then extensive training data collection and hand-crafted features are eliminated, but computational requirements increase
Solution Approach 1:
The neural networks are pre-trained offline before deployment, performing the computationally intensive training phase in advance when computational resources are available. During actual operation, the pre-trained networks require only inference, which is less computationally demanding than training, thereby simplifying deployment while managing energy consumption during field operation
Solution Approach 2:
The system uses pre-trained network weights and architectures that have been developed and optimized separately. These pre-trained models can be copied and deployed across multiple systems without requiring each system to perform its own extensive training, reducing both the computational burden and energy consumption during deployment while maintaining high performance
Data Source
Figure 1A
Figure 1B
Figure 1C
AI summary
The system and methods of the various embodiments enable better capturing of images of objects that are moving relative to a camera. A computing system may be configured to capture a frame containing a target object, generate at least two candidate object bounding boxes for the target object, and determine a candidate object feature vector for the part of the frame contained within each candidate object bounding box. The computing system may compute the distance between each candidate object feature vector and a target object feature vector, selecting the candidate object bounding box associated with the candidate object feature vector that is closest to the target object feature vector, generate partial image statistics on the part of the frame contained within the selected candidate object bounding box, and provide the partial image statistics as input to an AX algorithm that controls the camera.