Monocular Target Detection Model Using 2D Pixel Coordinates
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Monocular-based three-dimensional target detection methods are sensitive to external camera parameters and struggle with generalization across different scenes, due to reliance on three-dimensional coordinate information during model prediction, and face challenges in collecting and calibrating large training datasets.
Innovation Solution
A target detection method that uses a monocular vision system to extract two-dimensional pixel coordinates of a target's center point projection on an image plane, allowing for the prediction of three-dimensional positions without relying on external camera parameters, by training a multi-layer convolutional neural network to extract position information from training images, which includes pixel coordinates of the target's projection on the image plane.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If three-dimensional coordinate information is used during model prediction, then detection accuracy is improved, but sensitivity to external camera parameters increases and generalization across different scenes deteriorates
Solution Approach 1:
The patent extracts only the necessary two-dimensional pixel coordinates of the target center point projection from the three-dimensional coordinate information. By removing the dependency on external camera parameters while retaining the essential positional information needed for detection, the model achieves both accuracy and generalization across different scenes and camera installations.
Solution Approach 2:
The patent transforms the detection approach by working in the two-dimensional image plane dimension rather than relying on three-dimensional coordinate transformations. By predicting the two-dimensional pixel coordinates directly from training images, the model avoids the complexity of external parameter calibration while maintaining detection precision.
2Measurement precision
If external camera parameters are used for target detection, then detection precision is improved, but the complexity of data collection and calibration increases
Solution Approach 1:
The patent removes the requirement for external camera parameters from the detection system. By extracting and using only the two-dimensional pixel coordinates that can be directly obtained from images, the system eliminates the complex calibration process while maintaining detection precision through the learned relationship between image coordinates and target positions.
Solution Approach 2:
The model performs self-calibration by learning directly from training images without requiring external parameter input. The system serves itself by automatically adapting to different camera installations and scenes through the training process, eliminating the need for manual calibration procedures.
3Measurement precision
If three-dimensional coordinate information is relied upon, then position extraction accuracy is improved, but the difficulty of collecting and calibrating training datasets increases
Solution Approach 1:
The patent extracts only the essential two-dimensional pixel coordinate information from training data, removing the need for complex three-dimensional annotation and calibration data. This simplification makes data collection much easier while maintaining position extraction accuracy through the learned mapping from image coordinates to target positions.
Solution Approach 2:
The patent uses simple two-dimensional pixel coordinates as proxies for complex three-dimensional position information. These simpler, easier-to-obtain coordinates serve the same functional purpose for detection while being much cheaper and easier to collect in large quantities for training.
Data Source
AI summary
A target detection method, an electronic device, a roadside device and a cloud control platform are provided and relate to the technical field of intelligent transportation. The target detection method includes acquiring a detection image of a target to be detected, wherein the detection image is an image taken by a detection monocular camera for the target to be detected; extracting position information of the target to be detected by inputting the detection image into a target detection model, wherein the target detection model extracts the position information of the target to be detected from the detection image, and the position information of the target to be detected comprising pixel coordinates of a projection, on an image plane of the detection monocular camera, of a center point of an orthographic projection of the target to be detected on a ground where the target to be detected is located and determining a three-dimensional position of the target to be detected based on the position information of the target to be detected.