EARLY WARNING SYSTEM TO REDUCE MOTOR VEHICLES ACCIDENT CASES USING IMAGE RECOGNITION AND BAYESIAN NETWORK METHOD

• Main Author: Regina Aurora, Theophanie
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/64976
• Institution: Institut Teknologi Bandung

Insurance policies have been known to change driver behavior which is considered a moral hazard. These changes in behavior are usually in a negative sense in which drivers become less careful when driving, which in turn will increase the probability of an accident happening. As the probability of accident increases, the probability of claim also increases. Monitoring this change in behavior is inherently difficult. It causes asymmetric information between the insured and the insurer, as it can increase the probability of the insurance company’s losses. Therefore, this research will build a model which monitors driver behavior when driving and give an early warning based on certain behaviors in hopes to decrease the probability of accidents. The model requires image data in forms of people driving and the consequences of said behavior. For implementation purposes, a camera in the car is required to take pictures of the activities carried out by the driver while driving. The model will then take the photo which will be classified as a type of distraction with image recognition using a Convolutional Neural Network. The classification result will be accumulated and then the model will calculate the probability of an accident based on the result. The probability of an accident is evaluated using the Bayesian Network method. A K-Fold Cross Validation is used to choose the best parameter for each distraction. Furthermore, from the result of the calculated probability of accident, an early warning system will be designed to warn the driver if they are deemed to be distracted or in a mathematical sense, the driver’s probability of accident has exceeded a determined limit. From the research results for image recognition method, it is found that by using the AdaMax algorithm, the SoftMax function for hidden layers, and the ReLu function for output layer, the accuracy result is 99.70%, whereas if activation function in the hidden layers is changed to the SoftMax function, the accuracy will decrease to 55.05%. Moreover, if the algorithm used is changed to Adam with the same activation function as the initial function, an accuracy of 99.60% is obtained. From these results, the combination that gives the highest accuracy is the AdaMax algorithm, the ReLu function in the hidden layers, and the SoftMax function in the output layer. Meanwhile, obtained for calculating the chance of an accident using the Bayesian Network method using five distraction classifications can describe the chance of an accident well, that is the more distractions you do, the higher the probability value. Furthermore, by combining image recognition and Bayesian Network methods, an early warning system is obtained that has the potential to reduce the chance of accidents.