Image Depth Recognition Using Pose-Based Dynamic Object Separation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image depth recognition models struggle with poor training accuracy due to the movement of dynamic objects, leading to inaccurate depth information recognition in vehicle images.
Innovation Solution
A method involving the use of a depth recognition network and a pose network to separate static and dynamic objects, generate a pose absolute value matrix, and adjust the depth recognition model based on gradient and photometric errors to exclude the influence of moving objects, thereby improving accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If training images include both static and dynamic objects, then the training dataset is comprehensive, but the training accuracy deteriorates due to movement of dynamic objects
Solution Approach 1:
The patent segments the training images by separating static and dynamic objects through pose estimation and motion detection. By dividing the image into static regions and dynamic regions, the system can selectively process each type differently, thereby maintaining comprehensive training data while improving accuracy by excluding misleading dynamic object movements from depth estimation.
Solution Approach 2:
The patent extracts and removes the influence of dynamic objects from the training process. By identifying moving objects through pose networks and motion detection algorithms, the system extracts only the static background information for depth estimation training, effectively taking out the harmful dynamic elements that degrade training accuracy.
2Productivity
If the depth recognition model processes all objects in the image, then the processing is comprehensive, but the depth information accuracy deteriorates due to moving objects
Solution Approach 1:
The patent applies segmentation by dividing the image processing into separate streams: one for static objects and another for dynamic objects. The static object stream undergoes depth estimation while the dynamic object stream is either excluded or processed separately, ensuring comprehensive processing of all objects while maintaining high accuracy for static depth information.
Solution Approach 2:
The patent introduces pose estimation networks and motion detection algorithms as intermediary steps between image input and depth estimation. These intermediaries identify and flag dynamic objects, allowing the system to selectively process only static objects for depth information, thereby acting as a mediator that filters out harmful dynamic elements while maintaining comprehensive object awareness.
Data Source
AI summary
An image depth recognition method held in a storage medium and running in a disclosed electronic device acquires images to be recognized and two original images. An original image is recognized through a depth recognition network to obtain an initial depth image, and a pose absolute value matrix is generated based on the two processed original images and a pose network, the pose network and the initial depth image generating an initial projection image. The processed two original images are recognized according to the pose absolute value matrix and the preset threshold matrix and adjusted based on errors between the initial depth image, the target image, and the target projection image. The depth recognition network obtains a depth recognition model, and the depth information of the image can be recognized. The method can improve the accuracy of the depth recognition of the image.