Incident Prediction

Emergency Response Management (ERM) includes two major stepsof prediction of the demand (accidents) and allocation and dispatch of the resources(responders). Principled decision-making in ERM necessitates the use of accuratespatial-temporal prediction of accidents. By having an accurate model topredict the time and location of accidents, we can strategically allocate and dispatchthe resources.Traditional methods that aggregate past incidents over space(known as the hotspot method) are not helpful in making useful short-termpredictions when the spatial region is large. Additionally, If accidents are studiedin high spatial-temporal resolution, they are very rare events, which makes theaccident prediction a challenging problem. Consider the probability of accidentoccurrence in the closest street during the next 2 hours, which most likely isvery close to zero. Another aspect adding complexity to the problem is thelarge number of factors involving in accidents while most of them are unknownor unavailable. If not enough data is available, accidents can be considered random events. Our main goal is to design a spatiotemporal prediction model andpipeline using different techniques including statistical analysis, featureengineering, deep learning, graph theory, etc. An efficient machine learningmodel requires a lot of data. We collect various types of data, includingweather, traffic, highway geometry, historical accident records, etc. fromassorted resources. Depends on the goal of the project, other resources such as real-time crowdsourced Waze data can be used to increase the accuracy of the model.Our experiments have shown, applying this pipeline in the state of Tennessee witha total area of over 100,000 sq. km can reduce the response time by 10%.

A brief presentation of our work for SmartComp 2021 is availabe on our youtube channel. Please refer to the publication tab for more details. Here, please visit the poster about our work in the domain of emergency response was presented in the TDOT Innovation forum 2021. The study is also partially supported by the National Science Foundation under Grant No. 1814958. Additional funding and data were provided by Tennessee Department of Transportation (TDOT).