Self-Driving Cars: A Survey (1901.04407v2)

Abstract: We survey research on self-driving cars published in the literature focusing on autonomous cars developed since the DARPA challenges, which are equipped with an autonomy system that can be categorized as SAE level 3 or higher. The architecture of the autonomy system of self-driving cars is typically organized into the perception system and the decision-making system. The perception system is generally divided into many subsystems responsible for tasks such as self-driving-car localization, static obstacles mapping, moving obstacles detection and tracking, road mapping, traffic signalization detection and recognition, among others. The decision-making system is commonly partitioned as well into many subsystems responsible for tasks such as route planning, path planning, behavior selection, motion planning, and control. In this survey, we present the typical architecture of the autonomy system of self-driving cars. We also review research on relevant methods for perception and decision making. Furthermore, we present a detailed description of the architecture of the autonomy system of the self-driving car developed at the Universidade Federal do Esp\'irito Santo (UFES), named Intelligent Autonomous Robotics Automobile (IARA). Finally, we list prominent self-driving car research platforms developed by academia and technology companies, and reported in the media.

• The paper presents a comprehensive review of self-driving car architectures, detailing robust perception techniques such as LiDAR-based localization with 9cm accuracy.
• It examines decision-making systems that include efficient route and path planning, with methods like Transit Node Routing achieving microsecond-level query times.
• By segmenting system functions into perception and decision-making, the survey offers actionable insights for advancing SAE level 5 autonomy in autonomous vehicles.

Self-Driving Cars: A Survey

"Self-Driving Cars: A Survey" by Badue et al. offers a comprehensive review of autonomous vehicle research, explicitly focusing on developments since the DARPA challenges. The authors structure their examination around the typical architecture of self-driving car systems, dividing the discussion into perception and decision-making processes. This organization provides a clear delineation of the responsibilities and methods employed in each subsystem, from localization to high-level planning and control.

Perception System

The perception system is divided into several subsystems, each handling distinct tasks essential for an autonomous vehicle's operation:

1. Localization: Various methodologies for determining the vehicle's position without reliance on GPS are discussed. The paper compares LiDAR-based, LiDAR plus camera-based, and camera-based localization methods, detailing approaches such as Monte Carlo Localization (MCL) and Extended Kalman Filter (EKF). For instance, Siadat et al.'s approach achieved lateral and longitudinal errors of 9cm and 12cm, respectively, showcasing the precision of state-of-the-art LiDAR-based localization.
2. Mapping: Both offline and online mapping strategies are examined, emphasizing the creation of occupancy grid maps (OGMs) and other representations like OctoMaps and Gaussian Process Occupancy Maps (GPOMs). Offline mapping using GraphSLAM and online mapping using FastSLAM and other algorithms are crucial for developing comprehensive environmental models.
3. Road Mapping: Automated generation of road maps from aerial images or sensor data is critical. Techniques leveraging deep neural networks (DNNs) for extracting road features are highlighted, with examples such as a DNN achieving 83.7% accuracy in lane marking detection.
4. Moving Objects Tracking (MOT): The tracking and prediction of dynamic obstacles are dissected through traditional, model-based, stereo vision-based, grid map-based, sensor fusion-based, and deep learning-based methods. The integration of these methods allows for real-time tracking and collision avoidance.
5. Traffic Signalization Detection: The detection and recognition of traffic lights and signs utilize both model-based and learning-based approaches. Convolutional Neural Networks (CNNs) play a significant role, with systems achieving high accuracy rates, albeit with dependency on large, annotated datasets.

Decision-Making System

The decision-making system encompasses route planning, path planning, behavior selection, motion planning, and control:

1. Route Planning: Algorithms for finding the shortest path in road networks are scrutinized. Techniques like the A-star, Arc Flags, Contraction Hierarchies (CH), and Transit Node Routing (TNR) are evaluated for their efficiency and scalability, with TNR achieving query times as low as 2.09 µs.
2. Path Planning: Methods include graph search-based, interpolating curve-based, and sampling-based techniques. Variants of the A-star algorithm and spline curves are common, with practical applications shown in real-world autonomous vehicles.
3. Behavior Selection: The complexity of urban driving necessitates finite state machines (FSMs), Markov Decision Processes (MDPs), and ontologies to model and predict vehicle behavior in various scenarios. These approaches ensure adherence to traffic laws and safe navigation in dynamic environments.
4. Motion Planning: Generating smooth and safe trajectories involves numerical optimization and model-predictive control (MPC). These methods account for vehicle dynamics and environmental constraints, with the objective of generating feasible, collision-free paths.
5. Obstacle Avoidance: Real-time adjustments to planned trajectories to prevent collisions. Effective methods include dynamic obstacle anticipation and real-time velocity adjustments.

Case Study: IARA

The paper details the Intelligent Autonomous Robotic Automobile (IARA) developed at Universidade Federal do Espírito Santo, exemplifying the integration of discussed technologies. IARA's system architecture follows the typical self-driving car setup, achieving autonomous navigation of 74 km in urban settings.

Industry Snapshot

The paper lists key industry players and their contributions, such as Waymo, Uber, Tesla, Baidu, and Nvidia. These companies are crucial in driving forward the deployment and refinement of self-driving car technology.

Implications and Future Developments

The research outlined in the survey profoundly impacts both theoretical exploration and practical deployment of autonomous vehicles. Advancements in sensor fusion, machine learning, and real-time processing continue to push the boundaries. Future work will likely focus on achieving SAE level 5 autonomy, addressing ethical considerations, and enhancing the robustness of self-driving systems in diverse environments.