DRIVER BEHAVIOR PREDICTION USING FUZZY LOGIC HIDDEN MARKOV MODEL (FL-HMM) BASED ON OBJECT DETECTION AND TRACKING

• Main Author: Rizqullah Mahdi, Alif
• Format: Theses
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/68349
• Institution: Institut Teknologi Bandung

Driving is a daily process that most humans undertake as means of transportation. Recently, other than humans, autonomous systems are allowed to drive vehicles on the road. However, the main reason in autonomous vehicle accidents is caused by human driver maneuvers. Maneuvers by human drivers are directly influenced by the behavior of the driver. 93% of accidents that occurred in Indonesia are caused by human factors. Therefore, the ability to identify other driver behavior is crucial while driving. This research intends to help drivers, both humans and autonomous systems, to identify the behavior of other drivers surrounding them by utilizing a driver behavior prediction system. The system will assist drivers in anticipating the movement of other vehicles. Observational driving data is used as the basis for creating the prediction system. The prediction system proposed in this research utilizes Fuzzy Logic Hidden Markov Model (FL-HMM) for predicting driver behavior. FL-HMM is a combination of fuzzy logic and HMM. HMM is a concept that defines the changes of hidden states (immeasurable) based on observed states (measurable). Driver behavior HMM is modelled by using observed vehicle speed as observed state and lane selection as hidden state. Fuzzy logic is utilized to mimic the process that humans undergo to estimate object speed. The observed state or speed estimation is determined by comparing the observed vehicle speed with the speed of the observer. The HMM uses the output of the fuzzy logic to generate predicted hidden states. The FL-HMM is equipped with the Baum-Welch algorithm to update the HMM based on observational data. Moreover, the Viterbi algorithm is utilized to generate the driver behavior prediction. Predictions are determined based on vehicle detection and tracking with a combination of YOLOv5 and StrongSORT. The observed vehicle data are acquired by using GoPro HERO7 action camera. The prediction system is modelled by using two methods, i.e., analytical method and optimization method with Flower Pollination Algorithm (FPA). There are two prediction system created for each method, the original prediction system (OPS) and the modified prediction system (MPS). The prediction system is differentiated according to the assumption of driver behavior classification based on vehicle speed. The OPS utilizes two HMM to predict driver behavior in low and normal speed, whereas the MPS utilizes three HMM to predict behaviors in low, normal, and high speed. The resulting prediction systems, which are the OPS, MPS, OPS-FPA, and MPS-FPA, are tested to compare the performance of each model. Based on experimental results, the prediction system predicts driver behavior in low, normal, and high speed quite successfully. The best result for low speed prediction is 95.83% by OPS-FPA, for normal speed prediction is 59.80% by MPS-FPA, and for high speed prediction is 81.40% by MPS. Moreover, the overall best result is obtained by the MPS with an average prediction percentage score of 65.16%. The result shows that the predictions are preemptive therefore it can help drivers in anticipating oncoming dangers from other vehicles on the road.