A Survey of Autonomous Driving: Common Practices and Emerging Technologies (1906.05113v3)

Abstract: Automated driving systems (ADSs) promise a safe, comfortable and efficient driving experience. However, fatalities involving vehicles equipped with ADSs are on the rise. The full potential of ADSs cannot be realized unless the robustness of state-of-the-art improved further. This paper discusses unsolved problems and surveys the technical aspect of automated driving. Studies regarding present challenges, high-level system architectures, emerging methodologies and core functions: localization, mapping, perception, planning, and human machine interface, were thoroughly reviewed. Furthermore, the state-of-the-art was implemented on our own platform and various algorithms were compared in a real-world driving setting. The paper concludes with an overview of available datasets and tools for ADS development.

• The paper outlines core ADS challenges including increasing fatalities, regulatory issues, and complex operational design domains.
• The paper compares modular and end-to-end system architectures, examining trade-offs in sensor fusion and localization techniques.
• The paper highlights emerging sensor technologies and SLAM innovations that promise enhanced perception and decision-making in autonomous driving.

A Survey of Autonomous Driving: Common Practices and Emerging Technologies

The paper "A Survey of Autonomous Driving: Common Practices and Emerging Technologies" explores the current landscape and unresolved challenges in the field of Automated Driving Systems (ADSs). The authors, Ekim Yurtsever, Jacob Lambert, Alexander Carballo, and Kazuya Takeda, provide a comprehensive review touching upon the present hurdles, system architectures, and innovative methods in ADS development.

Core Discussion Points

1. Challenges and Implications of Autonomous Driving:
   • Fatalities and Trust Issues: Despite the promise of safety and efficiency, ADS fatalities are on the rise, posing significant barriers to the realization of ADS's full potential.
   • Operational Design Domains (ODDs): Level 3 automation and beyond remain challenging under diverse and unpredictable road conditions.
   • Regulatory and Societal Impact: ADS has profound implications on the mobility-impaired population and will potentially revolutionize concepts such as Mobility as a Service (MaaS).
2. System Architectures:
   • Modular vs. End-to-End Driving: Modular systems break down the task into localization, perception, planning, etc., whereas end-to-end systems directly map sensor inputs to driving tasks. The latter, despite its simplicity, lacks hard-coded safety measures, posing significant risks.
   • Ego-only vs. Connected Systems: The traditional ego-only systems maintain self-sufficiency, while connected systems leverage Vehicular Ad-hoc Networks (VANETs) to share data among vehicles and infrastructure. Despite their promise, connected systems face infrastructural and security challenges.
3. Sensor Modalities and Hardware:
   • Diverse Sensor Suite: The ADS relies on a combination of Lidar, radar, ultrasonic sensors, and cameras to ensure robust perception under varied conditions. The balance between higher accuracy (Lidar) and lower cost and size (radar) is crucial.
   • Innovative Sensors: Event cameras, despite their nascent stage, offer rapid response times potentially transforming dynamic object detection.
4. Localization and Mapping:
   • Fusion Techniques: GPS fused with IMUs provides a basic but insufficient solution for high-accuracy localization. State-of-the-art methods rely on pre-built maps and point cloud matching for greater accuracy.
   • Simultaneous Localization and Mapping (SLAM): While promising high accuracy without pre-built maps, SLAM implementations face computational efficiency challenges.
5. Perception Tasks:
   • Image-based and 3D Object Detection: The evolution from traditional methods to deep learning-based approaches drastically improves accuracy and efficiency, though challenges arise in dynamic lighting and weather conditions.
   • Semantic Segmentation: Segmenting road elements from camera imagery is crucial for tasks like lane detection. Deep convolutional networks have shown promise though real-time operationalization remains a hurdle.
6. Risk Assessment and Decision Making:
   • Uncertainty Quantification: Propagating uncertainty through Bayesian deep learning networks can significantly enhance the robustness of ADS.
   • Human Behavior Prediction: Predicting the behavior of surrounding human drivers remains an underexplored domain, crucial for long-term planning and safety.
7. Datasets and Open-Source Tools:
   • Annotated Datasets: Benchmarks like KITTI and recent datasets like nuScenes provide vital training and validation data for perception and planning tasks, fostering inter-disciplinary research and development.
   • Open-Source Frameworks: Tools like Autoware and CARLA simulator offer platforms for safe development and testing before real-world implementation.

Implications and Future Directions

The survey underscores significant strides in ADS technology while spotlighting pervasive challenges. Practical implications span regulatory, infrastructural, and societal domains. Theoretically, integrating human behavior models and enhancing multi-agent systems will propel the field forward. Future research will likely focus on:

• Improving Real-World Robustness: Addressing edge cases in perception and planning.
• Human-in-the-Loop Systems: Enhancing interaction between humans and ADS, especially during transitional phases.
• Cybersecurity: Ensuring secure V2X communications to prevent malfeasance.

This survey serves as an essential compendium for experienced researchers venturing into the multifaceted domain of autonomous driving, providing a clear picture of current practices, challenges, and emerging innovations.