EdgeNet: An End-to-End Deep Neural Network Pretrained with Synthetic Data for a Real-World Autonomous Driving Application
These videos have been matched automatically. Contact us if they are not relevant.
These videos have been matched automatically. Contact us if they are not relevant.
View abstract on PubMed
Summary
This summary is machine-generated.This study introduces an end-to-end autonomous driving architecture with edge detection. Integrating edge detection and synthetic data pretraining significantly improves performance and reduces training time for real-world applications.
Area Of Science
- Computer Vision
- Robotics
- Machine Learning
Background
- Autonomous driving systems require robust perception to handle diverse real-world conditions.
- Bridging the domain gap between synthetic training data and real-world scenarios is a significant challenge.
- End-to-end learning approaches offer a promising direction for autonomous vehicle control.
Purpose Of The Study
- To present a novel end-to-end architecture for autonomous driving incorporating edge detection.
- To address the domain gap between synthetic and real-world data for improved transfer learning.
- To evaluate the impact of custom edge detection layers and synthetic data pretraining on model performance.
Main Methods
- Developed an end-to-end architecture with custom edge detection layers preceding an Efficient Net module.
- Utilized RGB and depth images, along with inertial data (IMU), for training.
- Created a synthetic multimodal dataset encompassing 100 diverse weather and traffic scenarios for pretraining.
- Predicted driving speed and steering wheel angle.
Main Results
- The inclusion of edge detection layers enhanced transfer learning performance with both synthetic and real-world data.
- Pretraining the architecture with synthetic data led to reduced training times.
- Synthetic data pretraining improved the model's performance when applied to real-world driving data.
Conclusions
- Custom edge detection layers are beneficial for autonomous driving architectures, particularly for transfer learning.
- Synthetic data pretraining is an effective strategy to accelerate training and boost performance in real-world autonomous driving applications.
- The proposed architecture demonstrates a viable approach to bridge the sim-to-real gap in autonomous driving.
These videos have been matched automatically. Contact us if they are not relevant.
These videos have been matched automatically. Contact us if they are not relevant.
Related Concept Videos
Here, we present a method to obtain images of the intestine upon laser-induced wounding. By exposing the mouse intestine to a multiphoton laser, the loss of a single or multiple crypt(s) is induced locally. By repeatedly imaging the damaged area over months, the real-time dynamics of intestinal recovery is...
This article presents and describes an outpatient treatment for prostate cancer using focal laser ablation. Laser catheter placement is guided by MRI-ultrasound fusion imaging in a fashion similar to prostate needle biopsy. Treatment is monitored in real-time with a thermal probe, placed adjacent to the laser fiber.
The presence of cancer stem cells have been associated with relapse or poor outcomes after radiotherapy. This manuscript describes the methods to study the radiosensitivity of cancer stem cells in lung cancer cell...
During vacuum induction melting, laser-induced breakdown spectroscopy is used to perform real-time quantitative analysis of the main-ingredient elements of a molten...
An original experimental setup for heating cells in a culture dish using 1.94 µm continuous-wave laser radiation is introduced here. Using this method, the biological responses of retinal pigment epithelial (RPE) cells after different thermal exposures can be...
A protocol to cut neurons in C. elegans with a MicroPoint pulsed laser is presented. We describe setting up the system, immobilizing worms, and severing labeled neurons. Advantages include a relatively low-cost system and the ability to sever neuronal processes or ablate cells in...