Autonomous Vehicle 3D Object Recognition Without Real-Time Image Correction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing object recognition systems for autonomous vehicles are not robust to changes in camera parameters and suffer from excessive calculation and performance deterioration due to real-time correction processes.
Innovation Solution
An image-based 3D object recognition apparatus that includes a processor to train a network model by correcting distortion and rotation, generate distorted images, and use a feature transfer module to ensure consistent feature extraction, reducing the need for real-time correction and improving processing speed.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If real-time correction of distortion and rotation is performed based on camera parameters, then object recognition accuracy is improved, but calculation overhead increases and processing speed decreases
Solution Approach 1:
The system performs distortion and rotation correction during the offline training phase of the neural network model, rather than during real-time inference. The processor trains the model using corrected images that have had their distortion and rotation adjusted based on camera parameters. This preliminary action allows the model to learn from corrected data without requiring computationally intensive real-time correction during actual object recognition operations, thus maintaining high accuracy while improving processing speed.
2Measurement precision
If real-time correction of distortion and rotation is performed based on camera parameters, then object recognition accuracy is improved, but calculation overhead increases
Solution Approach 1:
The distortion and rotation correction operations are performed in advance during model training, not during real-time inference. The processor corrects images offline and uses these corrected images to train the neural network model. This transfers the computational burden to the training phase, allowing the model to incorporate correction transformations without requiring complex real-time calculation during deployment.
Solution Approach 2:
The system generates multiple distorted versions of corrected images during training by applying various distortion and rotation transformations. These transformed copies are used to train the model's robustness to different camera parameters and viewing angles, enabling the model to generalize better without requiring complex real-time corrections.
3Adaptability or versatility
If the system is made robust to camera parameter changes through real-time correction, then adaptability is improved, but processing time increases
Solution Approach 1:
The system performs robustness training in advance by training the neural network model on corrected images and multiple distorted versions during the offline training phase. This preliminary training allows the model to learn invariance to camera parameter changes without requiring time-consuming real-time corrections during inference, thus improving adaptability while minimizing processing time.
Solution Approach 2:
Multiple distorted copies of corrected images are generated during training with different distortion and rotation parameters. The model is trained on these diverse copies to learn robust feature extraction that is invariant to camera parameter variations, enabling the system to handle different camera configurations without additional processing time during deployment.
Data Source
AI summary
An object recognition apparatus and method for an autonomous vehicle are disclosed. The object recognition apparatus includes a processor and storage. The processor: trains a first network model based on an image; corrects distortion and rotation of the image; generates at least one distorted image based on the corrected image; trains a feature transfer module based on a feature extracted from the corrected image and a feature extracted from the at least one distorted image; inserts the feature transfer module into the first network model; and performs fine-tuning for a second network model including the feature transfer module based on the at least one distorted image.