Predictive Analytics for Traffic Management on I-270 in St. Louis

Safety Champion

Contact Photo
chief safety and operations officer becky allmeroth
Becky Allmeroth
Chief Safety and Operations Officer
Contact Info

Missouri Department of Transportation
105 W. Capitol Ave.
Jefferson City, MO 65102
Phone: (573)-751-2803
Bio: Link

Project Manager

Ploisongsaeng Intaratip
Senior Traffic Studies Specialist
Contact Info


Phone: (314) 315-7481


Purpose of the Project:

One of MoDOT’s major areas of focus is improving the safety of roadways. There are many factors that can cause crashes, ranging from driver behavior to roadway geometry to weather conditions. MoDOT has applied many different engineering and behavioral strategies in pursuit of safety, such as public outreach to educate younger drivers, promotion of safety campaigns, analyzing fatal and serious injury crashes to perform systemic safety improvements, and road safety audits. Despite these efforts, crashes still happen. Even with all of the roadway and crash data available to MoDOT it is difficult to predict future crash locations. MoDOT and other emergency responders must react to crashes as they happen and respond as quickly as possible from their current positions.

Predictive analytics is the integration of real-time and historical data sources into a single platform, frequently processed with the use of artificial intelligence or machine learning, for analysis and decision making in the near-term. An active construction project presents a challenge for traditional means of crash prediction due to the frequently changing roadway conditions. A predictive analytics engine can process and react to new data to quickly to spot trends, allowing it to identify the circumstances which can lead to crashes before they occur. MoDOT is the first DOT in the country to focus the use of this tool in a heavy construction area with the pilot implementation on the I-270 Project in the St. Louis District.

The MoDOT predictive analytics pilot will provide the ability to predict high crash risk areas up to 24 hours in the future. This will help MoDOT to better monitor potential high-risk crash areas and position its Emergency Response vehicles to be proactive in responding to incidents. It will also improve safety, resource efficiencies and response times to crashes.

Predictive Analytics – Starting in August 2021, the crash prediction and incident detection algorithms were evaluated for their accuracy every quarter. The crash prediction algorithm will be evaluated only for AM (7-10am) and PM (4-7pm) peak hours. The incident identification algorithm will be evaluated for fatal and serious injury crashes only. The results are based on crash data captured by Transportation Management Center operators in MoDOT’s Advanced Traffic System software, so it is likely that some crashes on Interstates went unrecorded and are not included in the data set.

The project scope was adjusted to get higher accuracy on both algorithms by including more connected vehicle data feeds, installing 74 static cameras to collect traffic data, and pursuing St. Louis County 911 Computer-Aided Design data. Please see below for the integration status of each data source:

  1. Wejo Connected Vehicle data – Live on March 7, 2022
  2. Traffic Vision – Live on April 18, 2022
  3. HAAS – Live on June 2022
  4. Volvo – Live on August 2, 2022
  5. iCone – Live on August 16, 2022
  6. 74 static cameras – Installed by September 9, 2022 and dashboard training on October 3, 2022
  7. Otonomo – Live on September 21, 2022
  8. St. Louis County CAD – in progress
  9. Surfsight – in progress
  10. Dash Cam – in progress

The first pie chart above, Percent of Crashes Predicted, shows the results for the visible and nonvisible crash risk location predictions during AM and PM peak hours for October. Visible crash risk locations are locations where the algorithm predicts an elevated risk of a crash occurring and displays them on the map. Nonvisible crash risk locations are locations where the algorithm predicts an elevated risk of a crash occurring and does not display them on the map. The crash risk location prediction algorithm can predict anywhere from 1 location to about 70 locations per 3-hour period. There were 3 locations based on highest crash risk probability shown for each period due to readability of the map. It negatively affects operator performance to show large numbers of locations at a time on the platform.

The second pie chart above, Percentage of Peak Hour Crashes Predicted, shows the percent accuracy of crash risk location prediction compared by month for the year of 2022. In order to evaluate the effectiveness of the algorithm, an analysis was done to identify what percentage of predicted crash risk locations experienced crashes during the prediction timeframe. In October, the results showed crashes in 4.9% of visible crash risk locations. However, there were crashes in 12.3% of nonvisible crash risk locations, where the algorithm predicted the correct locations of crashes but the locations were not visible on the map.

Based on the difference in percent accuracy between visible and nonvisible crash risk location predictions, MoDOT’s near term goal is to increase the percentage of crashes in visible locations to be the same or greater than the percentage in nonvisible locations. Therefore, MoDOT has changed the settings to show 5 crash risk location predictions per period instead of 3 locations. Evaluation of the new setting will be done in January 2023 to determine the effectiveness of the change.

MoDOT and Rekor are continuing to work together to improve of the accuracy of the results to meet the target accuracies of 10% of the total crashes predicted and 15% of the crashes greater than 1 hour prediction for visible crash risk location prediction.

Table 2 below shows the results for the incident identification algorithm for fatal and serious injury crashes. There were 313 total matching crashes between the ATMS and Rekor systems that could be used for the evaluation in October. The algorithm was rated against existing methods based on whether Rekor detected an incident prior to other tools including police radio, Waze, and CCTV monitoring, among others. October results showed a significant increase in incidents first identified by Rekor. In 7 months, Rekor was able to integrate 7 different data sources from 7 different vendors. Therefore, “incidents first identified by Rekor” showed an increase due to a significant amount of data feeds integrated into the platform.

Table 2: Incident Identification Algorithm Performance (Fatal and serious injury crashes only)



Incidents First Identified by Rekor

Incidents First Identified by Other

Incidents Identified Simultaneously

November 2021




February 2022




April 2022




July 2022 (249)

41.8% (104)

55.4% (138)

2.8% (7)

October 2022 (313)

54.1% (170)

42% (132)

3.8% (12)


Advanced Video Analytics – Traffic Vision was fully operating as of January 2022. October results showed a total of 981 incidents detected by Traffic Vision. 13 out of 981 incidents were pedestrians which is about 1% of the total incidents. The top 3 incident types are 340 stopped vehicle/object in the roadway (39%), 335 congestion (39%), and 179 slow speed (21%).

Table 3 shows the results for the advanced video analytics performance followed by example pictures of Traffic Vision alerts for multiple types of incidents. The percentage of false incidents and incidents which were unable to be verified for October decreased from July due to an improvement to the slow speeds setting. For more details on settings changes, please see challenges in the “Advanced Video Analytics” section below. The category “Unable to Verify” includes alerts that could not be verified due to speed of traffic, clarity of video, weather interference, or other conditions.

Table 3: Advanced Video Analytics Performance


True Incidents

False Incidents

Unable to Verify

February 2022




April 2022




July 2022 (975)

85% (855)

14% (107)

1% (13)

October 2022 (981)

88% (867)

12% (114)



Integrated Modeling for Road Condition Prediction (IMRCP)The platform was fully operational in February 2022. IMRCP is currently under implementation and will be used primarily for winter weather events. The project team has developed a winter weather event plan that will allow for evaluation of the platform. The project team is ready to engage this plan during the first winter weather event of the year. The results from the first event will be used to optimize the platform and the evaluation plan for additional winter weather events. The goal is to capture winter weather events with differing severities and intensities.