Detection Circuit Neural Network Distance Estimation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current distance detection devices in portable electronic devices face high production costs and inaccuracies due to factors like obstructions and lighting, while multi-stage models require significant memory and time, making them unsuitable for devices with limited performance.
Innovation Solution
A detection circuit with a neural network module and calculation circuit processes images from a single camera to generate an output tensor containing position and distance adjustment information, calculating an initial distance and adjusting it to estimate the distance between the camera and an object using a convolutional neural network (CNN) model.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If a single camera with deep learning model is used to predict distance, then device cost is reduced, but distance measurement precision deteriorates due to obstructions, light, and color factors
Solution Approach 1:
The patent segments the distance estimation task into two parts: (1) using a lightweight YOLOv5 model for object detection and initial depth prediction, and (2) using a separate ResNet-based refinement model to adjust the depth based on specific object characteristics. This segmentation allows each model to specialize in its function, improving overall precision while keeping the base system cost-effective.
Solution Approach 2:
The patent changes the parameter approach by introducing a depth adjustment factor that is dynamically calculated based on object type, lighting conditions, and occlusion status. Instead of using a single fixed depth prediction, the system adjusts the predicted depth parameter based on multiple contextual factors, thereby improving measurement precision without increasing hardware cost.
2Measurement precision
If multi-stage model prediction is used to improve distance accuracy, then measurement precision improves, but device complexity and calculation time increase significantly
Solution Approach 1:
The patent applies partial action by implementing a conditional multi-stage prediction system. The refinement stage using the ResNet model is only activated when necessary (e.g., when the object is partially occluded or lighting conditions are poor), rather than always executing the full multi-stage pipeline. This reduces average calculation time and device complexity while maintaining high accuracy when needed.
Solution Approach 2:
The system incorporates feedback mechanisms where the output of the first YOLOv5 model serves as input to the refinement model, and the adjusted depth information feeds back to modify subsequent detection results. This feedback loop allows the system to iteratively improve accuracy while managing complexity through selective activation based on feedback from environmental conditions.
3Measurement precision
If traditional time difference of arrival method is used for distance detection, then measurement precision can be achieved, but device cost and complexity increase due to multiple sensors required
Solution Approach 1:
The patent replaces the mechanical/optical sensor system (requiring multiple cameras or LiDAR) with a computational approach using a single camera and deep learning models. The YOLOv5 model processes images to estimate depth through learned features, substituting complex hardware systems with software-based algorithms that achieve comparable or superior precision using only a single camera.
Solution Approach 2:
The patent creates a virtual copy of the physical depth information by learning the relationship between 2D image pixels and 3D spatial coordinates through training data. Instead of directly measuring depth with physical sensors, the system copies depth information from labeled training images and applies it to new images through the trained neural network, thereby eliminating the need for complex sensor systems.
Data Source
AI summary
The present invention provides a detection circuit including a neural network module and a calculation circuit is disclosed. The neural network module is configured to receive an image to generate an output tensor, wherein the output tensor includes position information of a specific object and distance adjustment information. The calculation circuit is coupled to the neural network module, and is configured to calculate an initial distance between an image capture device and the specific object according to the position information of the specific object, and generate an estimated distance according to the initial distance and the distance adjustment information.