Abstract:As technology advances in almost all areas of life, many companies and researchers are working to develop fully autonomous vehicles. Such level 5 autonomous driving, unlike levels 0 to 4, is a driverless vehicle stage and so the leap from level 4 to level 5 autonomous driving requires much more research and experimentation. For autonomous vehicles to safely drive in complex environments, autonomous cars should ensure end-to-end delay deadlines of sensor systems and car-controlling algorithms including machine learning modules, which are known to be very computationally intensive. To address this issue, we propose a new framework, i.e., an environment-driven approach for autonomous cars. Specifically, we identify environmental factors that we cannot control at all, and controllable internal factors such as sensing frequency, image resolution, prediction rate, car speed, and so on. Then, we design an admission control module that allows us to control internal factors such as image resolution and detection period to determine whether given parameters are acceptable or not for supporting end-to-end deadlines in the current environmental scenario while maintaining the accuracy of autonomous driving. The proposed framework has been verified with an RC car and a simulator.
engineering, electrical & electronic,chemistry, analytical,instruments & instrumentation
What problem does this paper attempt to address?
The paper attempts to address the "last mile" problem between Level 4 autonomous driving and fully autonomous Level 5 driving. Specifically, the authors focus on how to ensure the safety and real-time performance of autonomous vehicles in complex and dynamic environments. To achieve this goal, they propose an environment-driven approach that dynamically adjusts internal parameters (such as sensor frequency, image resolution, prediction rate, and vehicle speed) to balance runtime accuracy and end-to-end latency.
### Background of the Paper
With the development of technology, many companies and researchers are striving to develop fully autonomous vehicles. According to the Society of Automotive Engineers (SAE) in the United States, Level 5 autonomous driving is different from Levels 0 to 4, representing a fully autonomous driving stage without driver intervention. However, the leap from Level 4 to Level 5 requires more research and experimentation, as Level 5 autonomous driving needs to ensure high accuracy and low latency in various complex environments.
### Main Contributions
1. **Identifying Real-Time Factors**: Through extensive experiments, the authors identified real-time factors affecting Level 5 autonomous driving, including environmental factors (such as weather, road conditions, etc.) and internal factors (such as sensor frequency, image resolution, prediction rate, and vehicle speed). They analyzed the impact of these factors in detail and measured their relationships and trade-offs.
2. **Proposing an Environment-Driven Solution**: Based on the above findings, the authors proposed an environment-driven solution that dynamically adjusts internal parameters to balance runtime accuracy and end-to-end latency. Specifically, they designed an admission control module that can set appropriate internal parameters based on the current environmental scenario to support end-to-end latency requirements while maintaining the accuracy of autonomous driving.
### Solution
1. **Real-Time Factors**:
- **Sensor Frequency**: Autonomous vehicles are usually equipped with various sensors, such as cameras, radars, and LiDARs. The working frequency of these sensors can be dynamically adjusted to adapt to different driving environments.
- **Image Resolution**: High-resolution images can improve the accuracy of autonomous driving but increase the burden on the CPU and GPU. Therefore, the minimum image resolution needs to be determined based on external factors (such as weather and urban complexity) and limit core utilization.
- **Prediction Rate**: The execution time of prediction operations varies with algorithm complexity and may also be affected by other factors (such as image resolution and CPU/GPU load). Therefore, increasing the prediction frequency is not always feasible or desirable.
- **Vehicle Speed**: High-speed vehicles require higher sensor and prediction frequencies to ensure motor control within a short time. To avoid missing deadlines, it may be necessary to limit vehicle speed or increase sensor and prediction frequencies, or reduce the execution time of required operations.
- **Weather**: Perception and sensing in adverse weather conditions are major issues. High-spec cameras and sensors can improve performance but may also increase execution time and response delay.
- **Environmental Complexity**: Depending on urban complexity (such as traffic volume, pedestrian density, traffic signals, and road maps), the required perception and prediction rates and accuracy may vary. These external environmental factors may lead to different or inaccurate outputs of perception and prediction operations.
2. **Proposed Solution**:
- **Environment-Driven Control Method**: First, identify uncontrollable environmental factors, then set internal factors (such as sensor frequency, image resolution, prediction rate, and vehicle speed). By considering the trade-off between runtime accuracy and latency, feasible deadlines and internal parameters of sensors and algorithms can be set.
- **Optimization Operations**: Use optimization operations to calculate multiple values or boundaries for each internal factor and adjust based on driving feedback. Specifically, optimization operations include objectives such as maximizing image resolution and minimizing total utilization.
### Experimental Evaluation
The authors conducted experiments using a real RC car test platform and the CARLA simulator. The experimental results show that the proposed environment-driven approach can effectively balance runtime accuracy and end-to-end latency, thereby improving the safety and real-time performance of autonomous vehicles.
### Conclusion
By identifying real-time factors and proposing an environment-driven solution, the paper addresses the "last mile" problem from Level 4 to Level 5 autonomous driving. The experimental results validate the effectiveness of this approach, providing new ideas and technical support for achieving fully autonomous driving.