CNN-LSTM architecture for predicting hazardous driving situations

Abstract

This study aims to investigate how a CNN-LSTM model can be used together with recorded vehicle data from trucks and external weather data in order to predict a hazardous driving situation. The dataset consists of three-second-long driving snippets from customer and development trucks registered within Europe. The combination of a CNN and LSTM was implemented using two different architectures, one parallel and one sequential. The models were compared to a Random Forest classifier, as well as to a CNN and an LSTM individually. All models were evaluated with the complete dataset, data without weather features, and noisy data. The results from the complete dataset revealed that the Random Forest classifier achieved the highest accuracy of 92%, followed by the parallel CNN-LSTM with an accuracy of 81%. All models except the Random Forest classifier performed better with noisy data. The outcome of the thesis challenges the initial hypothesis that a CNN-LSTM is the optimal model given the context.