Vehicle Misbehavior Prediction
SwRI is researching a new area of transportation to predict when a driven veicle is about to commit a traffic violation or behave in an otherwise unsafe manner. The main objective is to discover the precursors to unsafe behavior and develop a system that is able to alert either the driver of the misbehaving vehicle or the surrounding drivers before the unsafe action occurs.
Feature Extraction/Data Cleansing
By using statistical software such as R and RapidMiner, SwRI has the capability to evaluate large data sets of vehicle trajectory data in a variety of ways, including:
- Principle component analysis
- Latent semantic analysis
- Dynamic time warping
- Feature ranking
- Creation of new features from existing data
- Detection and removal of erroneous data
A common problem with large datasets is the discovery of samples of interest. Instead of finding a needle in a haystack, often the issue becomes finding a specific needle in a stack of needles. At SwRI we have developed techniques that allow us to comb through the data in an automated fashion to locate such outliers and anomalies.
SwRI has expertise in predictive analytics including:
- Clustering classification regression
- Time-series analysis and forecasting
- Classification and regression trees
- Rule induction algorithms
- Feature reduction techniques
- Parallel machine learning architectures
- Bayesian probabilistic models
- Multivariate analysis
- Data visualization
Research and Development Progress (to date)
Internally, SwRI performed preliminary analysis starting with the results of a previous traffic study that was available on the Next Generation Simulation (NGSIM) website. The study produced processed video clips of roadway segments that identified and classified individual vehicles and other associated parameters that were saved into trajectory files.
We took these parameters, saved them in a customized database, and performed extensive pre-processing to cleanse, transform, and analyze the original data for our objectives. For instance, this transformation involved the use of R and RapidMiner to convert the location and time information into other useful forms, such as lane position and velocity. Next, we applied some initial analytic techniques to classify the data.
Using this data and its resultant analysis, we were able to predict if a vehicle was going to exit the roadway or continue on. Further, we intend to continue analysis of this data to construct the classifiers that will be used as a training set to facilitate evaluation of subsequently gathered data. Part of this continued analysis goes beyond standard statistical and transform techniques, and explores distance and time series techniques, such as Minkowski Distances and Dynamic Time Warping.
Additional descriptions of our continued efforts and the associated results will be presented in the future. The Driver Misbehavior Prediction project is following a traditional predictive data analytics approach, exploring a variety of tools and models, and utilizing an iterative process as shown in the following graphic.
The results of this work will provide a robust data analytics approach to the detection of anomalous vehicle behaviors and a mechanism for using the patterns discovered to predict the likelihood of future safety violations. The processes and methods developed also should be applicable to the large data sets produced by connected vehicles using vehicle-to-vehicle (V2V) or vehicle-to-infrastructure (V2I) communications, as well as many other domains.