This article will continue to analyze the core technologies involved in autonomous driving for everyone. The content of the previous "Perception" can be viewed in the original historical text:
2. Decision-making
In a relatively mature autonomous driving technology system, if the environmental perception module is compared to human eyes and ears, then the decision planning module is equivalent to the brain of an autonomous vehicle.
When making decisions and planning, autonomous vehicles will obtain road topology information, real-time traffic information, obstacles (traffic participants) information, and status information of the host vehicle itself from the environmental perception module.
Combining the above information, the decision-making planning system will analyze the current environment and then issue instructions to the underlying control execution module. This process is the main task of the decision-making planning module. 
Autonomous vehicle architecture diagram (picture source: see reference 3)
In other words, the behavioral decision-making and path planning of autonomous vehicles refer to the planning of a given number of alternative safe paths based on environmental perception and navigation subsystem output information through some specific constraints, and selecting an optimal path as the vehicle to drive. The process of trajectory.
This article will introduce in detail the technical structure system, technical methods, mainstream algorithms and chips of the autonomous driving decision-making planning module.
Ξ 1. Technical structure system
The common technical structure system in the field of autonomous driving decision-making and planning can be divided into hierarchical, reactive, and hybrid.
1) Layered hierarchical architecture
The hierarchical hierarchy can be understood as a series structure, the various modules of the autonomous driving system are arranged in a straight line, and the content processed by the previous module will directly enter the next stage, as shown in the following figure: 
hierarchical hierarchy The advantage is that the order of each module is clear, and the progressive structure of each module gradually reduces the scope of work handled by each module, and the accuracy of processing problems gradually increases, making it easier to achieve high-level intelligent control.
However, there are some problems with the hierarchical structure. First of all, the hierarchical structure needs to call sensor information in real time, which has higher requirements for sensors. In addition, the hierarchical layout, from environmental perception to execution control, there is a certain delay in the middle, lacking real-time and flexibility.
Finally, the hierarchical series structure has the problem of low reliability. In junior high school physics, everyone has been exposed to the knowledge of series and parallel circuits. Compared with parallel mode, the biggest problem of series mode is that no part of the overall system can have problems, otherwise the transmission channel of information flow and control flow will be affected, and the whole system will be in a state of collapse at any time.
2) Reactive system structure
The biggest difference between the reactive architecture and the hierarchical architecture is that the reactive architecture uses a parallel structure, as shown in the following figure: 
In the reactive architecture, the content of the decision planning module is arranged in parallel mode, and the environment is aware The content of will be simultaneously transmitted to multiple decision-making planning modules, which can highlight the characteristics of "perception-action" and easily adapt to a completely unfamiliar environment.
Compared with the hierarchical system structure, the reactive structure system occupies less storage space, has fast response and high real-time performance. At the same time, the parallel structure improves the stability of the overall structure, and the failure of one level of content in the decision-making planning module will not affect the normal operation of other levels. However, this also increases the complexity of the overall system operation and requires the support of higher levels of intelligent technology.
3) Hybrid architecture of the two
Due to the existence of certain problems in both the hierarchical and reactive architectures, a single system cannot meet the actual needs of autonomous driving to handle complex and changeable scenarios. Therefore, the hybrid architecture has received more and more attention. 
The hybrid architecture combines the advantages of the two. Global planning and local planning are applicable to different architectures, making autonomous vehicles more adaptable to complex and changing real road conditions.
Ξ 2. Technical method
According to the degree of mastery of environmental information, autonomous driving path planning can be divided into global path planning and local path planning.
1) Global path planning
Global path planning can also be called driving task planning. The main content is the planning of the driving path range.
Global path planning will plan an ideal path for autonomous vehicles in a known environment. The accuracy of path planning depends on the accuracy of information obtained by the environment perception module. This is a kind of pre-planning, similar to the "navigation" function commonly used in our daily life: input the departure place and destination, and the APP will automatically plan an optimal route.
It should be noted that global path planning requires accurate information about the environment in advance. When the environment changes, the planning results are likely to be invalid.
At present, the commonly used global path planning algorithms include Dijkstra and A* algorithms, as well as their improved versions.
The Dijkstra (single source shortest path) algorithm was proposed by scientist Edsger W. Dijkstra in 1956, and is mainly used to solve the problem of finding the shortest path between nodes in a graph. 
Dijkstra algorithm operation process
The advantage of this algorithm is that the given path is optimal, but the disadvantage is also obvious, that is, the computational complexity is high, because it is exploring the surroundings and there is no clear direction.
The A* algorithm was published in 1968 by Peter Hart, Nils Nilsson and Bertram Raphael of the Stanford Research Institute and is considered an extension of Dijkstra's algorithm.
The essence of this algorithm is breadth-first search, which can find the target node as soon as possible by adding conditional control on the basis of breadth-first search. The formula is expressed as: f(n)=g(n)+h(n), f(n) is the evaluation function from the initial point through the weight node n to the target point, and g(n) is the value from the initial point in the state space The actual cost from node to node n, h(n) is the estimated cost of the best path from n to the target node.
2) Local path planning
Local path planning can also be called real-time path planning. In an environment with obstacles, self-driving cars will use their own sensors to perceive the surrounding environment in real time, find an optimal local driving path, avoid collisions and maintain a safe distance.
The advantage of local path planning is that it can feedback and correct the planning results in real time, ensuring that the autonomous vehicle is always in the optimal driving path. The disadvantage is the lack of global environment information, and it may happen that the correct path or the complete path cannot be found.
In general, there is no essential difference between global route planning and local route planning. The two work together, and autonomous vehicles can better plan the optimal path.
Ξ 3. Algorithm and Chip
Currently, there are mainly three types of behavior decision algorithms commonly used in autonomous vehicles:
1) Based on neural networks: The decision-making system of autonomous vehicles mainly uses neural networks to determine specific scenarios and make appropriate behavioral decisions.
2) Rule-based: Engineers come up with all possible combinations of "if-then rules", and then use rule-based technical routes to program the car's decision-making system.
3) Hybrid route: Combining the above two decision-making methods, optimized by centralized neural network, and perfected by "if-then rules". The hybrid route is the most popular technical route.
In the chip field, there are two main types of commonly used mainstream autopilot chips, one is the Mobileye® EyeQX™ series vehicle computing platform developed by Intel-Mobileye, and the other is the NVIDIA Drive PX series vehicle computing platform provided by NVIDIA.
Mobileye was established in 1999 with the main goal of developing and promoting visual assistance systems in the transportation field. It was acquired by Intel Corporation in 2017 and used it as the core of its autonomous driving technology research and development strategy.
Take Mobileye's latest EyeQ5 chip as an example. The computing performance of EyeQ5 reaches 12 Tera/sec, and it can support up to 20 external sensors (camera, radar or lidar).
In addition, EyeQ5 has a heterogeneous, fully programmable accelerator. The four types of accelerators built into the chip are all algorithmically optimized, which is more conducive to the realization of "sensor fusion." 
The NVIDIA DRIVE platform is an AI platform launched by NVIDIA for autonomous driving. It combines deep learning, sensor fusion and surround vision technology to change the driving experience.
The NVIDIA DRIVE platform can understand the changes in the environment around the vehicle in real time, accurately locate itself on a high-definition map, and plan a safe route ahead.
Currently, DRIVE PX2 is the latest version of this product. It is manufactured based on the 16nm FinFET process with a TDP of 250W, supports 12 camera inputs, laser positioning, radar and ultrasonic sensors, and has a single-precision computing capacity of 8 TFlops. It is currently an advanced autonomous driving platform.
To be continued...
References for this article:
1. "2017 China Artificial Intelligence Series White Paper-Intelligent Driving"-China Artificial Intelligence Society
2. "2018 Artificial Intelligence Autonomous Driving Research Report"-Aminer
3. "Autonomous Vehicle Decision and Control"-Tsinghua University