Autonomous driving literature study "A Survey of Autonomous Driving: Common Practices and Emerging Technologies"

This is a comprehensive paper on autonomous driving, published in April 2020.
The full text systematically introduces the development history and development status of the autonomous driving system, explains the difficulties and challenges currently faced by the autonomous driving system, and then proceeds from the autonomous driving system framework, sensor classification, map positioning, information perception and evaluation, path planning, etc. In this aspect, the advantages and disadvantages of existing research techniques and methods are analyzed and compared.

1 Prospects and challenges

For social impact, it can reduce traffic accidents, traffic jams, and exhaust emissions; reallocate driving time, promote logistics development, etc. The level of driving automation is divided into five levels. At present, high-level level 4 and level 5 cannot be achieved.

2 System architecture

The author discusses the overall framework of the autonomous driving system from the connection level and algorithm level :

• Connection level:
  One type is a single vehicle system, which always performs all necessary autonomous driving operations on a self-sufficient car without relying on other vehicles and infrastructure; the other type
  is the Internet of Vehicles system, which emphasizes the realization of vehicles through the network Information exchange and sharing.
• Algorithm level:
  One type is module division, which is divided into different modules according to different functions;
  the other type is an end-to-end system, which obtains operational information through sensor input. The main methods include directly supervised deep learning, neuroevolution and deep reinforcement learning. .

3 sensors

Hardware organizations can be classified into five categories:

• Exteroceptive sensors for perception
• Proprioceptive sensors for interior vehicle condition monitoring tasks
• communication array
• Actuator
• computing unit

Usually in the form of multi-sensors, radar or lidar is used to make up for the camera's shortcomings in depth information. Laser has higher accuracy within 200 meters, but is more susceptible to weather effects than radar. Radar has the advantages of long range and low cost.

4 Map positioning

1. GPS-IMU fusion: Correct accumulated errors at time intervals. The problem encountered is that in environments with poor signals such as high buildings and tunnels, GPS positioning errors will increase.
2. SLAM: No prior information about the environment is required, but the calculation requirements are high and the real-time performance is poor.
3. Prior map-based positioning: Compares real-time information with detailed pre-built map information to find the best matching location. The problem is how to update the map when the environment changes.

5 Environmental information perception

1. Image-based target detection: including static object recognition (traffic lights, traffic signs, road intersections), dynamic object recognition (vehicles, pedestrians and bicycles). Current state-of-the-art methods basically rely on DCNN, but there are currently clear differences between them:
   ■ Single-stage detection frameworks use a single network to generate object detection locations and category predictions simultaneously. The mainstream algorithms include: YOLO and SSD.
   ■ The region generation detection framework is divided into two distinct stages, first proposing general regions of interest and then classifying them through a separate network of classifiers.

2. Semantic segmentation:
   Only using rectangular frames to select targets is less effective. Instance segmentation is further used to distinguish objects with different trajectories and behaviors.
   Segmentation methods: Mask, R-CNN, DeepLab.

3. Three-dimensional target recognition:
   It mainly relies on point cloud data of point lidar, and extracts features for recognition through filtering and clustering.

4. Target tracking:
   Estimating the heading and speed of a dynamic target in order to apply a motion model to track the target and predict the future trajectory of the target. Tracking is usually performed using a sensor fusion strategy.
   ■ Filtering methods: Bayes filter, Kalman filter, particle filter, etc.
   ■ Deep learning: CNN, multi-task network, etc.

5. Lane line detection:
   ■ Drivable area detection: detects the current driving lane of the lane and can be applied to lane departure warning, lane keeping and adaptive cruise control.
   ■ Complex semantic understanding of roads: identify other lanes and directions, and merge road structure information such as turns.

6 evaluation

1. Risk and uncertainty assessment:
   Quantify the level of uncertainty and risk in driving scenarios using Bayesian methods, Hidden Markov Models (HMM) and language models.
2. Ambient driving behavior assessment:
   Autonomous driving decisions in real environments are related to the behavior of surrounding drivers. Use HMM to predict the future behavior of the target vehicle.
3. Driving style identification:
   principal component analysis method, GMM-based driver model, etc.

7 Planning and decision-making

• Global planning:
  -goal-directed
  -separator
  -based -hierarchical and
  -bounded-hop.

• Local planning:
  - graph-based planners
  - sampling-based planners
  - interpolating curve planners
  - numerical optimization.

8 Human-computer interaction

Communicating with the driver through the Human Machine Interface (HMI), the most basic of which is the interaction of driving tasks, and the Auditory User Interface (AUI) is one of the options.

The current problems are the impact of environmental noise and the understanding of some content.

9 Datasets and Tools

• Datasets: PASCAL VOC, KITTI, DeepDrive, etc.
• Open source frameworks: Autoware, Apollo, Nvidia DriveWorks, openpilot, etc.
• Simulator: CARLA, TORCS, Gazebo, SUMO, etc.