The Interaction Of Data Analysis And Road Safety In Autonomous Vehicles

Data has become the new oil to the Mobility Industry. We have been seeing a major change in the way people move from point A to point B with the advent of car sharing and ride sharing businesses like Uber, Lyft and Ola in the past decade. The ride sharing business is only going to explode when L4 [1] autonomous vehicles become possible, as it reduces the biggest operating cost i.e. driver. Operating a ride sharing service in a city with autonomous vehicles involves a massive data analysis exercise. Road Safety which is another major expected outcome of autonomous vehicles can be ensured and enhanced by looking at a variety of data generated out of these operations.

Autonomous vehicle data can be classified into two broad categories. One is the functional data i.e. the sensor and map data which is required for the autonomous driving system to function. The other is the operations data i.e. all the data related to AV fleet operations like the number of AVs deployed in a municipality/city, passenger occupancy, trip information, pickup, drop off locations etc.

An AV should be able to do the following functions , each of which consumes and generates data, which we referred to as functional data above.

Sense-Plan-Act: The vehicles machine learning algorithms should analyze the huge chunks of data to predict correct outcome.

Mapping: The AV’s computer must possess a highly detailed 3D map which comprises of street features such as street signs, curbs etc.

Sensing in real time: Using sensors such as cameras and LIDAR (Light Detection and Ranging) an AV should be able to create a short distance read out in real time.

Vehicle to Vehicle communication: This scenario is not part of any AV at present but by using IoT (Internet of Things) the AV’s can communicate among each other to share important data such as traffic signals or speed signs, allowing them to adjust their behavior accordingly.

Autonomous Vehicles use High definition 3D maps which are highly data heavy unlike the Google Maps that we human drivers use. Building and maintaining the 3D map of a neighborhood or a city involves Terabytes of data storage, handling and constant updates. Just one autonomous car can generate as high a 40 terabytes of data per day. Maps is one of the major examples of incoming data. It’s won’t be a one-time download, those maps should be extremely detailed and continuously updated which can be used for lane control and road hazards. The data collected from the lidars and cameras on the autonomous vehicles operating in a city can help in keeping these 3d maps updated. Consider the example of a blocked lane on a highway due to an accident. If an autonomous vehicle passes by the spot and updates the 3D map in real time, which is then used by other AVs in the fleet, all the other AVs regulate their speed, change lane ahead of time , thereby reducing the congestion and increasing safety.

The high bandwidth functional data i.e. the sensor and map data is generally stored in the in-vehicle computers of an autonomous vehicle, which is periodically transferred to cloud storage. Analysing this functional data of an AV fleet operating in a city can yield a lot of insights for increasing road safety of that city. Some of my ideas in that direction are these. The data can be analysed to detect locations in a city where accidents are more likely to happen and take necessary actions.

Example 1: Analyse perception data and identify the roads where fellow vehicles are operating in a higher unsafe speed. The city can be informed to increase speed policing in those areas.

Example 2: Analyse vehicle acceleration data and identify the roads where the jerks in ‘z’ direction of the AV are higher which can mean the road is in a bad condition in that area. The city can be informed to repair the road in those areas.

Example 3: Analyse perception and map data and identify the places where the location of barriers or dividers are changed. This can be extremely useful in identifying the places where barriers/divider positions are disturbed or encroached into a lane that could otherwise prove to be extremely dangerous.

In the most recent case of Tesla’s Autopilot accident in Mountain View, California in 2018 [6], it was revealed during further investigation that the crash attenuator, a highway safety barrier that reduces the impact of a crash was found missing. It was either removed or not replaced after a prior accident. Such mishaps can be avoided if the proposed AV data analytics framework is implemented to inform the city officials about such dangerous locations.

Another vital application of data analytics in the AV industry is proving its own technology i.e. proving the road safety worthiness of Autonomous Vehicle Technology. Road safety is a major basic expectation of autonomous driving and it will take millions of miles of data to prove that. Waymo which leads the pack of AV players in the industry tested its technology over 5 millions of miles[3] of road driving data and still has a lot of ground to cover before they release it to the general public significantly. Analysis of human interventions, categorizing the interventions, estimating the root cause of intervention and informing the appropriate engineering team within the organization is going to be an ongoing data analysis exercise within any AV company both in the current testing phase and in the future as they roll out new software upgrades.

In conclusion, autonomous vehicles will make the roads very safe but it’s going to be a few more years before that happens. For the AV’s to be at top of the chain in road safety, as mentioned above lots and lots of different categories of data should be collected and analyzed, i.e. millions of miles of test drive data, traffic patterns, road signs, weather conditions and their effect on particular roads for example hilly areas or highways through forests, human interventions.