Author | Jessie
Produced by | Yanzhi
Weilai Automobile's automatic driving crash last year once again refreshed people's understanding of smart cars, whether the party responsible for the accident is the owner, or the car manufacturer, which may not be a conclusion that can be directly drawn from the survey results, in fact, as far as the current status of the level of smart cars is concerned, the level of safety is not the most important, the most critical is how to clarify the responsibility for the accident. However, even if the responsibility is clearly identified, the impact of this matter is far more than that. Because even the driver's misoperation may be attributed to the lack of publicity for the use of the system, or the system does not carry out an effective takeover reminder on its design boundaries.
What is a safe smart car, and how people deal with cars with autonomous driving functions, these are all questions worth pondering. Because even if a certain model of its automatic driving control ability is in an extremely superior state, it can not be 100% accident-free, according to statistics, for autonomous vehicles, 2% of accidents are inevitable, 4% are caused by unknown reasons. On top of that, 24 percent of perceived accidents, while impaired driver abilities accounted for another 10 percent, and the latter two combined 34 percent could be avoided through autonomous driving. The other 60% rely on system settings and people's decisions and preferences.
Therefore, it can be said that there are still many unavoidable situations in autonomous driving, including: errors in judgment, such as misestimating the speed, direction change and driving clearance of another vehicle; planning errors, such as too fast or too slow; and execution errors, such as errors at the driving operation level. Then this kind of collision hazard that cannot be completely avoided in practice can only be divided through ex post facto supervision. Because, if the responsibility for the accident cannot be clarified, whether it is legal, moral or user recognition, it seriously hinders the development process of automatic driving.
The development of traditional driver assistance systems usually uses a number of fixed weight indicators such as safety, comfort and energy saving designed by established rules and human designs, and simulates the dynamic decision-making and control of the vehicle by simulating the driver's dynamic decision-making and control based on the premise of boundary working conditions. However, the development of autonomous driving systems should be based on scenarios without boundary restrictions, and the ability to cover complex environments and unfamiliar, sudden and other scenarios is the core issue of system development.
Based on this, from the perspective of the gradual development of the entire autonomous driving system, it is necessary to improve the corresponding development and construction capabilities from different aspects in order to avoid the limitations of its perceived planning response. At present, many OEMs or tier1 have begun to make large-scale layouts for this piece of capacity building, such as proposing several mainstream solutions to respond to the response, which can be summarized as follows:
1) Improve perception ability - lidar + high-precision positioning + vehicle-road coordination;
2) Training perception performance - shadow mode acquisition system;
3) Strengthen data logging - data recording system;
4) Realize post-event monitoring - autonomous vehicle monitoring system.
The first two are mainly based on the data collection developed by the front-end to improve the algorithm capability, and the latter two are process analysis and accountability through post-event monitoring.
Improved environmental awareness
The frequent accidents of NIO or Tesla vehicles are nothing more than two important reasons, one is that the millimeter-wave radar or camera in the current autonomous vehicle does not recognize the stationary target in the environment, or does not react at high speed to the stationary target in front of it (which is also the main reason for the current accidents in Weilai cars); the other is because most of the accidents that rely on visual detection at night or under strong light tend to cause blindness (most of the accidents in Tesla vehicles are caused by visual limitations).
The former can also solve the problem of detection of stationary targets through lidar, because the scanning of lidar can target any different types of obstacles in front, and even cover multiple special-shaped vehicles, falling rocks, traffic accident vehicles and other scenarios, so in the case of low and medium speed, it can be a good solution to the detection and collision avoidance problems of stationary targets in front. In addition, lidar can also solve the collision problem caused by the temporary entry of close targets, and through simulation tests, we found that the perception scene with lidar can be improved by more than 50% compared with the traditional Radar+Camera.
But even so, it is still powerless when approaching the front car at high speed, because even if the corresponding stationary front car is not pinned, it is still not guaranteed to be able to slow down and brake in the current high-speed state of the vehicle. Here we can give an extreme example:
For example, the target distance of the conventional car in front of the current long-distance lidar can be recognized as 300m, if it is a general motorcycle, it will be less than 200m, if it is 0-120km/h according to the speed of automatic driving activation, approaching the front car at the highest speed, at this speed, it is easy to produce two different safety hazards:
First, when the vehicle recognizes the target in order to avoid collision, it will brake sharply with an average deceleration of about 2.77m/s2, which is easy to cause greater discomfort to the driver or passenger in the car;
Second, when the vehicle cannot recognize the stationary target in time, or when the time of recognition of the stationary target is too late, it will cause a collision with the environmental target in front;
Based on this, it is particularly critical to join the vehicle-road collaborative capacity building. Our sight anticipates dangerous targets ahead, which can greatly increase the identification distance to provide protection for timely and effective braking. Of course, the current generation of automatic driving can not really realize the car road or vehicle information communication.
The latter one due to insufficient light at night caused by inaccurate target detection problem, there are two solutions, the current use of self-driving activation of automatic headlights to alleviate, to be honest, this can only solve the close distance of the lane information detection, usually, this distance detection will not exceed 150m, this headlight state for the detection of lane information is only a drop in the bucket. In order to cure this type of problem, performance improvement may only be achieved by means of a night vision system.
Vehicle data acquisition training system
In order to cope with the shortcomings of the current automatic driving system in environmental cognition and data training, some OEMs and suppliers often tend to use more acquisition equipment to develop effective environmental acquisition and analysis algorithms to obtain data models in the scene training library. At present, the vast majority of R & D institutions in the process of testing the automatic driving system, on the one hand, found that the automatic driving algorithm can not cover the current working conditions, testers rely on manual recording of the problem; on the other hand, with the deployment of automatic driving data collection solutions, Tesla as an example began to analyze the real user's characteristic data during the driving process to improve and upgrade the current automatic driving algorithm.
1. Shadow mode operation principle
The development and establishment of test data sets for autonomous driving systems need to rely on numerous data sources and data domains to ensure that autonomous driving can achieve the necessary level of safety. At the same time, by mapping to a simulation system that simulates the driver's driving style, the bionic optimization of the automatic driving system can be realized.
The overall system architecture of the shadow mode is represented as follows, including collecting big data, simulating and analyzing the vehicle-side data, trigger recording, cloud management, simulation reconstruction, and rapid testing of algorithms. It involves a large number of scene analysis to establish feature-level data semantic expressions; collect driver control signals for driver behavior rationality analysis; collect driving data from body feedback to achieve control command difference analysis; based on the behavior difference data-mechanism analysis to achieve the classification triggering and control of learning networks; and finally use certain simulation training algorithms to locate environmental perception problems and decision planning problems.
So can the real shadow mode really solve most of the data collection, annotation and algorithm training problems? The author believes that there is still some uncertainty.
2, the short board of shadow mode
The core of "shadow mode" is that in the manned state, the system including sensors is still running but is not involved in vehicle control, but only verifies the decision algorithm. And the process requires that shadow mode can provide more and larger extreme conditions, including labeled and non-labeled training data. Shadow mode, on the other hand, relies on the human driver's driving decisions to annotate and trigger data backhaul. Therefore, the whole process of running the shadow mode algorithm still needs to ensure that the autopilot chip has high computing power. At the same time, the trigger mechanism is to trigger the recorded behavior of the system based on the criterion of driver control as reasonable, and this judgment standard is not necessarily correct, such as the scene of deceleration, the driver does step on the accelerator or hit the steering. In addition, the evaluation mechanism of the shadow mode is not scientific, because whether the algorithm of automatic driving is accurate or not needs to be located in the four sub-modules of specific perception, planning, decision-making, and control, and the final manifestation is often only one. This makes it impossible to accurately identify which submodule is the problem in the subsequent data analysis. And the shadow mode collection vehicles are not necessarily autonomous vehicles, it may be auxiliary driving vehicles, facing the L2 level, whether such assisted driving vehicles can be used for L3 levels is uncertain. Because the sensor of data acquisition, the recording algorithm triggered by the regulation module are inconsistent.
Based on this, even if many developers advocate equipping the entire shadow mode package on the self-driving model, it may not be able to really make up for the model's annotation and detection learning needs for almost all scenes.
Autonomous Vehicle Monitoring and Accountability System
Engaged in automatic driving, in fact, no one is willing to sacrifice, let alone to the machine or algorithm to sacrifice. Then the core problem becomes, once there is an accident, what happens in the whole process, why it will lead to the accident, the need for a clear record and playback process, in the process of test verification and even follow-up road test, everything that occurs can be traced back within a certain range. At present, the more pertinent is the automatic driving data recording system, which is also a system that the National Standards Committee requires the next generation of autonomous driving functional vehicles to be strongly marked. However, it is worth mentioning that the automatic driving data record only records the relevant information of the vehicle itself, the driver's operation, and the surrounding vehicles in the entire process of the safety accident from the perspective of the vehicle end. Our developers need to know more to better analyze vehicle data, not only for accident scenarios, but also for scene modeling and data calibration in development scenarios.
In particular, on April 7, 2021, the Ministry of Industry and Information Technology issued the Guidelines for the Administration of Access to Intelligent Connected Vehicle Production Departments and Products (Trial Implementation) (Draft for Solicitation of Comments), which puts forward the following requirements for the entry and listing of L3 or above autonomous driving functional models:
"It is required to establish and improve the enterprise safety monitoring service platform, ensure product quality and production consistency, monitor the operating status of (each test) vehicle, and record the driving trajectory, control mode, vehicle movement status parameters, driver and human-computer interaction status, driving environment information, vehicle actuator control information, takeover information and other data."
For automobile manufacturers, it is also hoped that through the monitoring platform, the monitoring of the operating state of autonomous vehicles can be realized, and through the recording of autonomous vehicle data, user behavior analysis, vehicle fault analysis and statistical analysis of the operating status of the automatic driving system, it is also hoped that the continuous iterative optimization of products, fault analysis, accident responsibility, etc. will be supported.
At present, a considerable number of enterprises in the operation of autonomous vehicles have established a platform for monitoring the operation of vehicles. For example, Changan, GAC, China Automobile Research Institute and other enterprises are developing monitoring platforms for new energy vehicles: real-time monitoring of vehicle status, trajectory, battery power supply, etc.; statistical analysis of fleet operation, historical trajectory playback, etc.; and real-time alarm analysis of battery electric drive system. Whether it is new energy monitoring or autonomous vehicle monitoring, it is essentially the use of vehicle networking data for in-depth value mining, from the dimensions of vehicle operation monitoring, analysis and fault warning, etc., to improve the management of vehicles and the use of vehicle network data.
The following is the functional division of the entire monitoring platform, as shown in the following figure, the monitoring platform is mainly used for the business level of the operation function and status analysis of the autonomous vehicle, and the statistical analysis of the operating state or frequency of the autonomous vehicle. Among them, there are 6 major business chains from bottom to top, which are the underlying function management, the basic function management, the statistical analysis management, the data analysis management, the platform docking management, and the top-level monitoring management. From the perspective of accident analysis, the whole process can be seamlessly monitored to collect the status of autonomous driving information in the vehicle, and it seems indispensable for the later division of responsibility.
summary
At present, whether from the national planning, enterprise capacity building, we hope that in 2025 can achieve automatic driving in a certain sense in a certain sense, at least within the scope of its ODD, can drivers completely not take over the vehicle, or even pay attention to the driving situation of the vehicle, but is it really OK? The author is conservative. Is it bicycle intelligence or vehicle-to-road collaboration, or both? Is the sensing end a high-cost lidar or pure vision technology, or hybrid sensing? These are still being discussed.
At present, whether it is Tesla or Weilai on the road of automatic driving, the pit is enough to make us wake up, automatic driving still has a long way to go, of course, there are some OEMs in the edge of the ball, such as proposing to first achieve automatic assisted driving function, but this also needs to provide better user instructions and human-computer interaction instructions to avoid misuse and abuse. The series of measures mentioned in this article for the possible failure of autonomous vehicles can enhance confidence in the development and application of autonomous driving to a certain extent, and hope that the road to automatic driving is still a road that is recognized and trusted by people.