Accelerate the implementation of intelligent driving projects? You still lack a truth evaluation system

Intelligent driving system has evolved rapidly , Not only does it need to greatly improve the perception ability of the system , We need to strictly check the real level of the perception system .

Can be observed , At present, the number of sensors on new models released by major automobile enterprises is often as high as dozens , Including lidar , camera 、 Millimeter wave radar 、 Ultrasonic radar, etc , There are more and more kinds of sensors .

But the improvement of sensing system capability is not simply hardware stacking , The evaluation of the real ability of the perception system will be the most critical line of defense .

It means , We have to be clear about , Whether the sensing system or even a single sensor can accurately identify all kinds of obstacles in complex scenes ？ More accurately , How should car companies verify the real effect of the perception system in complex driving scenes ？

The solution given by the mature automobile industry is “ Effective evaluation ”.

Of course, the evaluation work should use artificial intelligence

Obviously , Traditional evaluation methods have been difficult to meet the needs of intelligent vehicle development .

The perception system of intelligent driving is a software and hardware system based on artificial intelligence and big data , The traditional artificial scene test can not cover the real scene Coner Case, And the rental fee is expensive （ Tens of thousands / God ）.

In the industry , The intelligent driving function adapts to the development of a new scene , It requires the support of truth value data collected and marked by tens of thousands or even hundreds of thousands of real road test frames . meanwhile , With the increasing demand for intelligent driving application scenario support , The demand for road test truth data is growing exponentially , Will achieve PB Level , Is the “ big data ”.

On the ground , In a matter of 7 Days of road test scenario data , Just marking requires several annotators , It takes months of continuous work to complete . And if the traditional manual frame by frame annotation method is adopted , The error caused by judging the shape and size of obstacles in the annotation results is completely uncontrollable .

therefore , Just mark one item , The cost of traditional evaluation methods 、 Efficiency and accuracy , When facing the evaluation project of intelligent driving multi scenario big data , There's no way to parry .

（ Traditional manual annotation , Low efficiency 、 Great error 、 The high cost ）

But you know , According to the automobile industry V Model development process , Testing and development go hand in hand , Every key point in the development process , Must pass the complete evaluation verification . in other words , Only the implementation of verification is fast and accurate , Product development can be promoted rapidly .

This is also the reason why evaluation and verification must pay attention to efficiency and quality .

Since the perception system is an artificial intelligence system , Evaluation also faces “ big data ”, We also need to use artificial intelligence to solve evaluation problems .

The marking efficiency is thousands of times higher than that of manual work The truth data is reliable and abundant

that , How to use artificial intelligence to solve evaluation problems ？

The answer given by Suteng juchuang, who has been deeply engaged in the field of lidar sensing technology for more than ten years, is ：RS-Reference Truth generation and evaluation solutions .

According to introducing ,RS-Reference It has two parts , Truth generation system And Perception evaluation system .

To ensure the quality of real data ,RS-Reference Make full use of artificial intelligence technology , The truth generation system adopts multi-sensor fusion acquisition system + High performance offline AI Perception Software , The former assists in collecting raw data , The latter intelligent operation outputs truth value data .

RS-Reference High standard design starts with the original data acquisition system , It uses 128 Line lidar 、 Special lidar for blinding 、 Long range millimeter wave radar 、 Industrial cameras 、 Optical fiber IMU, High precision positioning RTK And other multi-sensor fusion schemes , Effectively ensure the redundancy and accuracy of the original data .

（RS-Reference Original data acquisition sensor scheme ）

RS-Reference Core high performance offline AI Perception Software , Integrated RoboSense More than ten years' accumulation of patented technology of point cloud perception , And cutting-edge high-performance artificial intelligence algorithm technology , use Offline processing mechanism Track and detect obstacles throughout their life cycle , Excellent perceptual performance .

When the test vehicle equipped with the sensor fusion scheme completes data acquisition , The truth generation software will automatically label the data , With It is thousands of times more efficient than manual marking , Output truth information .

Benefit from advanced high-performance offline AI Perception Software , The truth value information is accurate, reliable and rich in types , Including the speed of obstacles 、 Location 、 size 、 Categories and ID And other dynamic information , It also includes lane lines 、 Road boundary 、 Static information such as road curvature .

（ offline AI Annotation technology , High labeling efficiency 、 High precision , Truth value is rich in information ）

Full chain evaluation tool , Accelerate development validation cycle iterations

RS-Reference Of Perception evaluation system Some can provide customized evaluation and complete tool chain , To meet the needs of different customers .

Its rich software functions include data management 、 Data visualization 、 Manual calibration 、 perception KPI Report generation, etc . meanwhile , Provide for cameras 、 Laser radar 、 Millimeter wave radar and other sensor evaluation tools , And support customized sensor evaluation , Meet the requirements of automobile enterprises 、tier1 And other different types of sensors .

（ Intersection scene , Finished visualizing the truth value of the dimension ）

RS-Reference The test validation program also covers data cleaning 、 Data alignment 、 Data Association 、 Truth value matching, etc , Go deep into program optimization 、 The process of cyclic verification of adjustment , Greatly improve the development efficiency of vehicle sensing system software .

adopt RS-Reference Obtain evaluation analysis report , Engineers will address the deficiencies of the perception system , Perform upgrade iteration , Then evaluate again , In this cycle , Until we get a most satisfactory perception scheme .

（ The parallel development mode of R & D, design, test and verification of projects in the automotive industry ）

Upstream and downstream of the enabling industrial chain The truth value evaluation system has great potential

From the perspective of coordinated development of upstream and downstream of the industrial chain , Intelligent truth evaluation system , It can be deeply adapted to car enterprises 、ADAS Solution provider 、 Third party evaluation organizations and other enterprises' pain points , Realize the improvement of development efficiency or business value .

For car companies , because Coner case The scarcity of , Autonomous vehicle need hundreds of millions or even hundreds of billions of kilometers of test mileage to be effectively verified .

adopt RS-Reference Truth system , Car companies 、ADAS In the early stage of project development, the manufacturer can run through the scene database data as far as possible , Optimize the auto drive system efficiently , Improve the safety performance of auto drive system .

meanwhile , Enterprise cars can also pass RS-Reference Evaluation Verify the supplier's sensor or sensing system solutions , Quickly complete supplier screening .

about ADAS For the solution provider , Not only can I use RS-Reference Improve the development efficiency of camera or millimeter wave radar sensing scheme , It is possible to RS-Reference Evaluate as a third-party tool , Objectively verify the performance of the self perception scheme , thus Win more recognition and orders from car enterprise customers .

For third-party testing agencies , It can be based directly on RS-Reference Provide evaluation service for perception system .

（ carrying RS-Reference Test vehicle for ）

according to the understanding of , Suteng juchuang RS-Reference The system has been launched 4 More years , Has served dozens of companies around the world OEM and Tier 1 The customer's extensive test fleet .

Based on rich experience , Suteng juchuang and Guoqi （ Beijing ） Intelligent network Automobile Research Institute Co., Ltd 、 Beijing Huawei Digital Technology Co., Ltd. and Xiangyang Da'an Automobile Testing Center Co., Ltd. jointly take the lead in the formulation of 《 Requirements and methods for perception evaluation of intelligent networked vehicle lidar 》, Put forward a set of perfect evaluation requirements 、 Method and index system , Make specifications and guidelines for the test of the front mounted vehicle mounted sensing system of the laser radar .

Predictably, , The wide application of intelligent truth evaluation system , It will not only become the consensus of the industry , Become an important guarantee for guarding the safety of intelligent driving , It will also become a model of close cooperation between human beings and artificial intelligence .