A probabilistic framework for identifying safety critical events in naturalistic driving time series data

One of the core types of analysis performed in naturalistic driving studies (NDS) is event based analysis (EBA). EBA considers surrogate events, called safety critical events (SCE), for crashes, since actual crashes are rare. In order to find potential safety critical events, constant parameter thresholds have been commonly used (Simons-Morton et al., 2011). A key challenge of NDS is determining whether these SCEs are safety-relevant requiring time-consuming and expensive manual video coding. In order to lower the time needed for manual coding, a new approach of identifying SCE is presented on the data of the UDRIVE NDS. In this approach, SCE triggers are defined based on the likelihood of certain events countering the inherited high number of false-alarms of constant thresholds (Simons-Morton et al., 2011). The approach provides a functional relationship between the threshold parameters (e.g. longitudinal acceleration and situational parameters such as speed). This function is based on the joint probability density distribution (JPDD) of the involved trigger parameters. The first step is to estimate the JPDD from a representative sample. Because the linearly binned kernel density estimate approach (Wand, 1994) is computationally fast and allows for estimating two dimensional distributions, it was used for estimating the JPDD (Deng, 2011). The trigger function is computed as the percentile line of the estimated JPDD by computing the cumulative probability function. This curve represents a dynamically changing threshold of trigger parameters depending on a chosen set of situational parameters such as velocity. This approach allows for determining events occurring with a specific probability. A polynomial fit can help determining an easy-to-implement representation of the trigger function. This approach has two advantages: Specific values of a constant trigger threshold do not need to be chosen, but instead values are the result of a general parameter setting. Furthermore, it provides a dynamically changing threshold for different scenarios, such as a lower threshold on longitudinal deceleration values for driving on a highway (higher speed levels) compared to urban driving (lower speed levels). We expect that this framework closes the gap between discrete forms of SCE triggers and helps to lower the number of false-alarm detections and therefore relieves the burden of manual inspection