back to overview
26
.
02
.
2021

Understanding driving behaviour

A quick journey behind the scenes
Matteo Maffetti
Data Scientist at Swiss Re

Technology and insurance

The technological innovations of the past few decades have radically changed our lives: helping us with our daily tasks and providing new means to access both the services we're interested in and the information we seek. Devices like smartphones and smartwatches are probably the first examples that come to mind when we think about high-tech products – accessible devices retailing at relatively low prices which are used by the majority of people and have made big changes to the way we approach the world. One of the key advantages of such technological devices is that they provide direct access to services, thereby benefiting the consumer and strengthening their relationship with service providers.

Insurance companies are important service providers that contribute to society by providing a financial claims process to assist in the event of detrimental incidents. This has been the main goal of insurers for a long time. However, given that claims payments logically act as post-loss measures, risk management measures could offer support in preventing losses from happening at all. It goes without saying that we can't plan for every eventuality in life and there are certain events that simply cannot be avoided, but understanding risks gives insurers the opportunity to sensitise individuals to behaviours that should be avoided, in line with the paradigm "prevention is better than cure".

Thanks to technological advancement, it is now possible to collect concrete data that can be used to study risky behaviours and enable industries to develop plug and play solutions for their customers. As explained in this blog by Movingdots, Swiss Re's technology hub in the automotive and mobility space, enabling intuitive customer experience to coexist with complex IT architectures is essential to ensuring success.

Swiss Re's telematics app, Coloride, has been developed by applying the paradigm of prevention outlined above to the field of motor insurance. Coloride can be easily installed on a smartphone and is able to analyse driving behaviours and raise awareness about dangerous driving conducts. In the next section we'll explain how our app helps drivers prevent risks, by focusing on Coloride's detection module, which is developed to spot dangerous driving habits like harsh maneuvers and phone distractions.

Building blocks for analysing driving behaviour

A core component of Coloride is how it detects habits which contribute to the complex task of characterising customers' driving behaviour, specifically maneuvers and phone distractions.  The diagram below summarises the general schema of the computation workflow.

Blocks for analysing driving behaviour


Coloride can identify several maneuvers, such acceleration, braking, cornering, harsh steering, driving around roundabouts and at junctions, U-turns, from raw accelerometers and GPS signals collected directly from phones' embedded sensors.

Raw signals are then elaborated by using standard and novel signal-processing methods in order to remove noise and transform them into useful information that describes the car's motions. Swiss Re's Milan-based Advanced Analytics team has designed algorithms in order to align phone accelerometer signals to the car reference frame so that they can be used in describing the vehicle motion dynamics. This step is quite delicate because drivers are not asked to put the mobile phone in any type of holder to keep it stationary in the car so the algorithm must distinguish between all the rotations that happen during the trip and apply the proper corrections to estimate the vehicle frame orientation. In addition, GPS signals are enriched with contextual map content delivered by HERE and used to detect and characterise maneuvers e.g. harsh braking in front of pedestrian crossing or near a school zone.

The detection module works in three main steps and processes data in parallel across different maneuvrer types (see Figure 2 for a visual reference). Firstly, potential maneuvers are extracted from the signals of GPS-derived features (acceleration, braking, cornering), accelerometer-derived features (harsh steering) and contextually-based features (roundabout, junction, U-turn). Once a potential event is extracted, the GPS-, contextual- and accelerometer-derived features are computed to characterise it as part of the second step. These features are used by machine learning models in the ruleset evaluation step to predict if a potential maneuver can be released in output to the third and final step. This evaluation step is necessary to remove potential non-harsh events, as the previous step can generate false positives, and its aim is to generate candidates that comprehensively cover all possibilities of actual harsh events i.e. prioritising recall performance instead of precision.

Machine learning models used to filter out "bad" candidates in the ruleset evaluation step are trained against data collected during both on-track and on-the-road car tests. Whilst the former is used to learn signal patterns characteristic of each maneuver type, the latter's purpose is to tune the models on realistic car trips. For that, we have equipped vehicles with professional devices that can collect high-quality accelerometer and GPS data, in order to understand the typical noise that affects phone signals vs high-device signals and to develop proper techniques to tackle the problem.

In the final stage, called post-processing, maneuvers from the previous steps are put together in order to produce a unified and coherent output, exploiting both disambiguation and contextual map-based validation logics.

Figure 2: The diagram illustrates the detection module steps for identifying harsh events.
Figure 2: The diagram illustrates the detection module steps for identifying harsh events.


The detection of phone distractions, which is also part of the detection module, is based on events available in the smartphones' operating system and on events and features that are computed starting from the accelerometer's signals. The first kind of events (the locking and unlocking of a phone) are used as possible start and end triggers of phone distractions. They are considered in combination with phone handling events computed from the accelerometer's signals. The two sources of information are used in order to detect phone usage whilst driving, e.g. reading notifications, texting and calling.

Figure 3: Insurer's dashboard insights.
Figure 3: Insurer's dashboard insights.
Figure 4: Harsh events displayed on the Coloride app's map.
Figure 4: Harsh events displayed on the Coloride app's map.


The output of the detection module is displayed on Coloride's map once a trip is completed. Together with speeding and contextual information, they form the building blocks of our Advanced Scoring services, which aim to evaluate the behaviour of the policyholders behind the steering wheel from an insurance perspective, with the scope of supporting them in mitigating risk during trips (behaviour steering) and providing insurers with valuable insights on their customers, to enable risk-based pricing. As presented in this telematics publication on the Italian telematics experience, risk-based pricing and behaviour steering are two of the five telematics value-creation levers, together with risk selection (performed, for example, in try-before-you-buy propositions), value-added services (like cross-selling of other insurance products, such as travel insurance once you drive through the border) and loss control (including first notice of loss and other loss mitigation measures).
Insurers can implement their telematics programs focusing on a single lever or a mix of them, according to their portfolio, their innovation appetite and their reference market. Swiss Re's Automotive & Mobility Solutions and Movingdots are ready to support in all cases. If we have triggered your interest, get in touch with us to schedule a demo.

Share Article