AI-based framework for automatically extracting high-low features from NDS data to understand driver behavior
Metadata[+] Show full item record
Our ability to detect and characterize unsafe driving behaviors in naturalistic driving environments and associate them with road crashes will be a significant step toward developing effective crash countermeasures. Due to some limitations, researchers have not yet fully achieved the stated goal of characterizing unsafe driving behaviors. These limitations include, but are not limited to, the high cost of data collection and the manual processes required to extract information from NDS data. In light of this limitations, the primary objective of this study is to develop an artificial intelligence (AI) framework for automatically extracting high-low features from the NDS dataset to explain driver behavior using a low-cost data collection method. The author proposed three novel objectives for achieving the study's objective in light of the identified research gaps. Initially, the study develops a low-cost data acquisition system for gathering data on naturalistic driving. Second, the study develops a framework that automatically extracts high- to low-level features, such as vehicle density, turning movements, and lane changes, from the data collected by the developed data acquisition system. Thirdly, the study extracted information from the NDS data to gain a better understanding of people's car-following behavior and other driving behaviors in order to develop countermeasures for traffic safety through data collection and analysis. The first objective of this study is to develop a multifunctional smartphone application for collecting NDS data. Three major modules comprised the designed app: a front-end user interface module, a sensor module, and a backend module. The front-end, which is also the application's user interface, was created to provide a streamlined view that exposed the application's key features via a tab bar controller. This allows us to compartmentalize the application's critical components into separate views. The backend module provides computational resources that can be used to accelerate front-end query responses. Google Firebase powered the backend of the developed application. The sensor modules included CoreMotion, CoreLocation, and AVKit. CoreMotion collects motion and environmental data from the onboard hardware of iOS devices, including accelerometers, gyroscopes, pedometers, magnetometers, and barometers. In contrast, CoreLocation determines the altitude, orientation, and geographical location of a device, as well as its position relative to an adjacent iBeacon device. The AVKit finally provides a high-level interface for video content playback. To achieve objective two, we formulated the problem as both a classification and time-series segmentation problem. This is due to the fact that the majority of existing driver maneuver detection methods formulate the problem as a pure classification problem, assuming a discretized input signal with known start and end locations for each event or segment. In practice, however, vehicle telemetry data used for detecting driver maneuvers are continuous; thus, a fully automated driver maneuver detection system should incorporate solutions for both time series segmentation and classification. The five stages of our proposed methodology are as follows: 1) data preprocessing, 2) segmentation of events, 3) machine learning classification, 4) heuristics classification, and 5) frame-by-frame video annotation. The result of the study indicates that the gyroscope reading is an exceptional parameter for extracting driving events, as its accuracy was consistent across all four models developed. The study reveals that the Energy Maximization Algorithm's accuracy ranges from 56.80 percent (left lane change) to 85.20 percent (right lane change) (lane-keeping) All four models developed had comparable accuracies to studies that used similar models. The 1D-CNN model had the highest accuracy (98.99 percent), followed by the LSTM model (97.75 percent), the RF model (97.71 percent), and the SVM model (97.65 percent). To serve as a ground truth, continuous signal data was annotated. In addition, the proposed method outperformed the fixed time window approach. The study analyzed the overall pipeline's accuracy by penalizing the F1 scores of the ML models with the EMA's duration score. The pipeline's accuracy ranged between 56.8 percent and 85.0 percent overall. The ultimate goal of this study was to extract variables from naturalistic driving videos that would facilitate an understanding of driver behavior in a naturalistic driving environment. To achieve this objective, three sub-goals were established. First, we developed a framework for extracting features pertinent to comprehending the behavior of natural-environment drivers. Using the extracted features, we then analyzed the car-following behaviors of various demographic groups. Thirdly, using a machine learning algorithm, we modeled the acceleration of both the ego-vehicle and the leading vehicle. Younger drivers are more likely to be aggressive, according to the findings of this study. In addition, the study revealed that drivers tend to accelerate when the distance between them and the vehicle in front of them is substantial. Lastly, compared to younger drivers, elderly motorists maintain a significantly larger following distance. This study's results have numerous safety implications. First, the analysis of the driving behavior of different demographic groups will enable safety engineers to develop the most effective crash countermeasures by enhancing their understanding of the driving styles of different demographic groups and the causes of collisions. Second, the models developed to predict the acceleration of both the ego-vehicle and the leading vehicle will provide enough information to explain the behavior of the ego-driver.