Examiner: Bengt Arne Sjöqvist, Dept of Electrical Engineering
Supervisor: Stefan Candefjord, Dept of Electrical Engineering
The absence of a vehicle shell around the motorcyclist makes the motorcycle crashes fatal as the rider collides directly with the objects. During motorcycle crashes, when the rider is driving alone, there is no one to call for emergency rescue and first aid. Moreover, when the rider is in the countryside or at a time or place where there is no one to help, there rises a need for an automatic emergency alert system for motorcycle riders. This report is about the thesis with Detecht motorcycle crash detection company. Developing a crash prediction algorithm that can be used with Detecht’s smartphone-based application. It includes the crash prediction using a machine learning algorithm in python. Using the parameters from the sensors that are built-in in all smartphones nowadays. Eliminating the need for any separate device required for collecting the data needed to predict a crash and generate an alert message accordingly.
The crash samples in Detecht’s data that includes the 26 simulated crashes and 3 real crashes are a small number of samples as compared to the 940214 samples of normal driving samples recorded from 76 Detecht’s smartphone application users. It makes this crash prediction, an anomaly detection case. To have the best accuracy in results and to reduce the false alarms generated by the algorithm, different reasons for crashes are studied. These reasons show that a crash is not only a motorcyclist’s mistake but there are many other factors involved in them. Convolutional Autoencoder is the machine learning algorithm used for this anomaly detection, which is an artificial neural network. The algorithm works by reducing the dimensions of the data and regenerating the data using the learned reduce dimensions. Mean error distribution is used to evaluate the output of the algorithm by comparing the algorithm output to the actual data. This comparison predicts the crashes out of normal driving of the motorcyclists.
The report consists of a comparison of the convolutional autoencoder results with the previously used algorithm by Detecht and discuss problems using the machine learning algorithm to predict crashes. The behavior of the convolutional autoencoder is studied by relating the algorithm response to the respective sensor values at the time of the crash. A total of 85% of the 940214 data samples are predicted correctly having no false alarms and 62 % of the 26 simulated crashes were predicted correctly using the convolutional autoencoder. The results of the thesis emphasize more on reducing the false alarms from the algorithm. Since these false alarms can be irritating for the user which will result in users not trusting the application.
Keywords: Crash prediction, motorcycle crashes, anomaly detecion, machine learning, convolutional autoencoder, python.
Student project presentation
24 February, 2021, 14:00
24 February, 2021, 15:00