Exploring the Values of ADAS and DMS Data

From 2015 to 2017, I worked for the first aftermarket ADAS startup in China. In November 2017, I officially joined Momenta, a leading autonomous driving technology company. I saw a clear industry trend at that time, and later my judgment proved to be correct. Today I’d like to share my recent thoughts on the values of ADAS and DMS data.

Before I start, let me make a simple definition: ADAS is the abbreviation of Advanced Driver Assistance Systems, but here we only discuss the forward camera-based collision avoidance system; DMS, Driver Monitoring Systems, mainly refers to camera-based driver fatigue, distraction, other abnormal driver behavior, and FaceID detection. Of course, you may have heard that Jiangsu Province of China also refers to DMS as DSM, which is short for Driver State Monitoring.

When people mention ADAS, the first name that comes to mind is the Israel company Mobileye; likewise, when people mention DMS, the Australia company Seeing Machines rings in the heads. However, I believe that the global popularity of such products will definitely have a great relationship with the China market, with more choices and higher cost performance. The following is the developing history of the ADAS and DMS industries in China:

• Since 2013, some Chinese startups have entered this industry.
• In 2016, this industry reached the peak of financing.

But before 2017, there were no successful cases of commercialization in this industry.

• At the end of 2017, two major projects in Shenzhen and Dongying were executed:

In October 2017, China Pacific Insurance (Group) Co. Ltd., together with its risk management partner, installed more than 5,000 sets of ADAS and DMS devices for its commercial auto insurance clients in Shenzhen.

In December 2017, The Dongying Transportation Bureau forced the local 11,000 dangerous goods transport vehicle to install ADAS and DMS devices.?

These two projects have greatly boosted the confidence and direction of startups in this industry. More and more ADAS or DMS startups are beginning to apply 4G connected ADAS and DMS devices to the fleet management, helping fleets reduce accidents and save lives. This solution helps the fleets reduce insurance costs and increase ROI. Therefore, many insurance companies are also venturing into this field.

However, due to the fact the total amount of device installed is still a small number and the total time of use is too short, these domestic ADAS or DMS manufacturers, TSPs, insurance companies, research institutions and government agencies are still not able to give valuable data reports to prove the benefit of ADAS and DMS systems in fleet management field. According to my understanding, only automakers and insurance companies have the ability to give detailed, precise and relevant data.?

Back to the topic today, I think the values of ADAS and DMS data have the following three dimensions worthy of deep exploration:?

Topic 1: Algorithm Gets Better Through Deep Learning

4G connected ADAS and DMS devices can upload two types of data: regular alarms and corner cases. Regular alarms, including types of events, pictures and videos, are used for fleet drivers’ monitoring and driving behavior analysis. Corner cases, meaning uncommon scenarios that confuses computers; the system will upload such data, including the types, pictures, and videos, to the cloud to improve the deep-learning algorithm accuracy. These corner cases can be considered the fuel for the AI engine: more corner cases could result in higher accuracy. Therefore, deep-learning based algorithms will get better as the number of users and driving distances increase. Of course, we need to get more abundant samples of data to make the perception algorithm more robust in varied conditions of roads, weather and vehicles.

Compared to the traditional algorithms, deep-learning algorithms have more potential. People generally believe that deep-learning can help startups surpass Mobileye and Seeing Machines.

Topic 2: In-depth Analysis of Driver Behavior Based on Visual Perception Data

In the past ten years, the driving behavior analysis is based on hardware, such as GPS tracker, T-box, dashcam, and MDVR. You may also know that Zendrive, many western insurance companies and some Chinese companies rely on mobile-based App to analyze driving behavior. Yes, it’s just another hardware carrier. The data used in traditional driving behavior analysis relies mainly on events generated by GPS, G-sensor and gyroscopes, as follows:

• Driving time
• Early morning, afternoon, or other tiring driving time
• Hard braking, harsh deceleration, and sharp cornering
• Overspeed (normal speed limit)?
• Overspeed when making a turn or going uphill and downhill
• Idling
• Geo-fence (make sure the driver drives according to the planned path)?

However, the driving behavior rating reports based on the traditional data and events are not always acceptable, and sometimes the score is not even directly related to the driver's driving behavior. The main issues are as follows:

• An experienced driver may hit brake to avoid any potential risks, but this kind of correct driving behaviors is considered bad by some driving behavior analysis models.
• Sometimes, the driver may make a sharp turn to avoid hitting pedestrians, which will also be considered bad.?

In contrast, driving behavior analysis based on ADAS and DMS is more accurate and can really help drivers improve their driving behavior.

• If a driver triggers more FCW, UFCW, HMW and PCW alarm events than others within the month, this proves that the driver's awareness of keeping a safety distance needs to be improved.
• LDW alarm events indicate that drivers often change lanes without turning lights, which is very dangerous for drivers, passengers and surrounding vehicles.
• Fatigue driving, distracted driving and making a phone call while driving are the main causes of accidents.
• Combining the traditional emergency events and overspeeding alarms, we can create a driving behavior analysis model based on artificial intelligence. This model can make a reasonable judgment for each event before the accident occurs.
• Compared to the traditional driving behavior analysis, the pictures and videos of the events are more acceptable.?

Through the active safety warning during the daily driving, the driver will get used to keeping a safe distance and turning the lights when changing the lanes. When an alarm is triggered, the driver will subconsciously pay attention to his driving behavior. In summary, traditional driving behavior data combined with ADAS and DMS perception data can create a more reliable and accurate driving behavior analysis model.

Topic 3: Is Driving Data From CAN, ADAS and DMS Valuable for Autonomous Driving?Simulation?

At this stage, most autonomous driving teams are only in the phase of using basic environment perception algorithms and high-precision mapping for basic path planning and driving decisions. Next, the simulation and real road test are the key to test the team's autopilot technology. But if you want to bring safety and driving experience to a higher level, deep-learning is the key.

• The videos of before and after the FCW/UFCW/HMW events help the autonomous vehicle keep a safe distance. I believe that the blind spots to humans are the same to autonomous vehicles. Through deep learning and massive data, the autonomous vehicle will perform better than humans in the same scenarios.?
• Before the L5 autonomous driving is realized, a human back-up driver is still necessary. DMS can be used to determine who’s in control. The DMS will always detect the driver's attention. If the vehicle is not driving itself, it is necessary to ensure that the human driver can take over quickly.
• Videos of vehicle collisions can be used as a simulation scenario to help the autonomous vehicle understand the cause and effect of the accidents. It is very meaningful to understand the road conditions and the driving behavior of the autonomous vehicle. Effectively reducing accidents.

That’s all I have for now. Please share with me your thoughts. Happy to discuss.