Source-linked AI summary
A Machine Learning Approach for Driver Identification Based on CAN-BUS Sensor Data
Md. Abbas Ali Khan, Mphammad Hanif Ali, AKM Fazlul Haque, Md. Tarek Habib
TL;DR
Prior driver-identification studies use telemetric and CAN-BUS data, with reported results spanning different driver counts and sensing setups. The study applies supervised classifiers to selected and window-processed vehicle data, using public OBD-II PIDs. With 15 features and two drivers from the Ocslab dataset, kNN achieved 99.99% accuracy (highest), while Naïve Bayes achieved 96.15% (lowest).
Problem
Prior driver-identification studies use telemetric and CAN-BUS data, with reported results spanning different driver counts and sensing setups.
Source: §2 · p. 2Method
The study applies supervised classifiers to selected and window-processed vehicle data, using public OBD-II PIDs.
Source: §4.3 · pp. 7–8, §7 · p. 14Results
With 15 features and two drivers from the Ocslab dataset, kNN achieved 99.99% accuracy (highest), while Naïve Bayes achieved 96.15% (lowest).
Source: §5 · p. 13Takeaways & Limitations
In the Ocslab evaluation using public OBD-II service 01, the reported algorithms provide their best accuracy with fewer drivers.
Source: §5 · p. 13The authors could not build a model using the full Ocslab dataset and report successful modeling for a partial dataset with two drivers.
Source: §7 · p. 14
Abstract
from arXiv · showhide
Driver identification is a momentous field of modern decorated vehicles in the controller area network (CAN-BUS) perspective. Many conventional systems are used to identify the driver. One step ahead, most of the researchers use sensor data of CAN-BUS but there are some difficulties because of the variation of the protocol of different models of vehicle. Our aim is to identify the driver through supervised learning algorithms based on driving behavior analysis. To determine the driver, a driver verification technique is proposed that evaluate driving pattern using the measurement of CAN sensor data. In this paper on-board diagnostic (OBD-II) is used to capture the data from the CAN-BUS sensor and the sensors are listed under SAE J1979 statement. According to the service of OBD-II, drive identification is possible. However, we have gained two types of accuracy on a complete data set with 10 drivers and a partial data set with two drivers. The accuracy is good with less number of drivers compared to the higher number of drivers. We have achieved statistically significant results in terms of accuracy in contrast to the baseline algorithm
2. Literature Review
The section reviews earlier driver-identification approaches and describes the OBD-II dataset, preprocessing pipeline, classifiers, and evaluation metrics used in this study.
2. Literature Review
- Prior work uses telemetric GMM/HMM, OBD-II CAN classifiers, and driving-style categories, with reported results varying across driver counts and feature sets.Examples include GMM accuracy of 81% for 12 drivers and 73% for 30, HMM accuracy of 85% for about 20, and CAN-BUS accuracy of 99% among 15 drivers.Source: §2 · pp. 2–3
4.1 Dataset Preparation
- The Ocslab dataset contains 94,401 records, 51 features, and trips from 10 drivers across city roads, freeways, and a parking lot.The data were collected from KIA vehicles in South Korea; the authors note that OBD-II cannot provide some signals, including body-control, airbag, and wheel-angle status.Source: §4.1 · p. 4
4.2 Data Prepossessing
- The preprocessing pipeline applies feature selection, normalization, and sliding-window processing to the 51-feature dataset before classification.The passages identify these steps but do not specify the full processing procedure here.Source: §4.2 · p. 5
4.2.1 Feature Selection
- Feature selection reduced the original 51 features to 15 using CorrelationAttributeEval and Ranker, with 10-fold cross-validation and seed 1.The authors describe removing homogeneous, irrelevant, superfluous, and mostly correlated features; engine torque and its correction feature are identified as identical.Source: §4.2.1 · p. 6
4.2.2 Data Normalization
- Min-max normalization integrates features on different scales, which the authors say is essential for algorithms such as kNN and SVM.The cited passage defines the minimum and maximum as the feature’s respective lowest and highest values.Source: §4.2.2 · p. 7
4.3 Sliding Window
- The sliding-window method shifts an overlapping fixed-length subwindow across the time series, calculating mean, median, and standard deviation for the 15 selected features.The authors describe computing these statistics from each subwindow.Source: §4.3 · p. 7
4.3 Description of the classifiers
- The supervised classifier suite includes kNN, SVM, Logistic Regression, Naïve Bayes, and REP Tree, using distance-, margin-, probability-, and pruning-based approaches.The passages describe kNN’s Euclidean distance, SVM’s margin-maximizing hyperplane, Naïve Bayes’ posterior probability, and REP Tree’s pruning procedure.Source: §4.3 · p. 8
4.4 Performance Metrics
- Performance is evaluated with accuracy, precision, F-measure, and recall from confusion metrics; the paper also describes averaging class confusion matrices for a final 2 × 2 matrix.The passages explain that false positives, false negatives, true positives, and true negatives are used to calculate these metrics.Source: §4.4 · p. 9
5. Results and Discussion
The results report high accuracy for two-driver classification, while the paper concludes that accuracy is better with fewer drivers. Comparisons with a baseline and a full-dataset analysis further qualify the reported performance.
5. Results and Discussion
- 99.99% accuracy was highest for kNN, while Naive Bayes was lowest at 96.15%, using 15 features and two drivers.The results were evaluated with 10-fold cross-validation, using nine blocks for training and one for testing in each fold.Source: §5 · p. 13
- The authors report that k-NN accuracy exceeds the 14.03% ZeroR baseline and describe the model as statistically significant.The full-dataset baseline analysis concerns drivers A and D.Source: §5 · p. 13
- The authors conclude that state-of-the-art algorithms provide the best accuracy with fewer drivers in the Ocslab dataset using public OBD-II service 01.Source: §5 · p. 13
- 29.93% accuracy for Naive Bayes on the full Ocslab dataset was compared with 14.03% for ZeroR.The passage also reports driver D accuracy increasing from 0.14% to 0.8%.Source: §5 · p. 13
- With kNN batch size 100 and Euclidean distance, accuracy declined as the nearest-neighbor count increased; REP Tree reached 99.95% for drivers A and D.The reported REP Tree size was 1397, with depth 26 and learning rate 0.001.Source: §5 · p. 13
5.1 Comparative Performance Analysis
- A comparative performance analysis is presented in Table 8.Source: Table 8 · p. 13
7. Conclusions
The reported accuracy is 99.99% for two drivers but 76.36 % for the full 10-driver dataset, which the authors consider unsuitable for driver identification. The conclusions also report limits in building a model from the full Ocslab dataset and describe classifier and computational trade-offs.
7. Conclusions
- 99.99% kNN accuracy among two drivers with 15 features contrasts with 76.36 % accuracy on the 10-driver full dataset, which the authors call unsuitable for identification.The authors say driver identification needs around 100 % accuracy.Source: §7 · p. 14
- The authors could not build a model from the full Ocslab dataset despite 24 hours of continuous processing, but successfully modeled a partial dataset with two drivers.The research uses only public PIDs; some non-public OBD-II PIDs are hard to identify.Source: §7 · p. 14
- Higher kNN neighbor counts require more processing time and yield less accuracy, while lower counts produced higher accuracy with less computational complexity.The authors report kNN accuracy of 99.99% among two drivers with 15 features.Source: §7 · p. 14
- The ensemble and other supervised learning classifiers achieved almost the same accuracy, and the authors report statistically significant results relative to a smaller baseline classifier.Source: §7 · p. 14