Automatic dangerous driving detection can contribute to improve traffic safety and prevent accidents. This work proposes a machine learning (ML)-based system for dangerous driving classification, using inertial sensor data collected from a self-driving car in a simulated environment. Dataset used in the experiments includes accelerometer and gyroscope readings to allow identify patterns associated with both dangerous and safe driving, which would be unfeasible under real-world conditions. Three ML classifiers were implemented and compared: SVM, FNN, and LSTM. Pre-processing techniques, such as filter decimation and moving average, were applied to optimize the data. Our results indicated that the LSTM model achieved the best performance due to its ability to handle temporal sequences, followed by SVM, which showed high precision and recall with filtered data. The FNN model demonstrated sensitivity to pre-processing, showing inferior performance without filtering techniques. It is concluded that the evaluated models have great potential for practical applications, such as alert systems, contributing to accident reduction, and promoting safer traffic.