This repository contains all the implementations (including proposed models as well as baselines) and sample files for the D-CRNN paper.

• Statistical Feature Matrix: to generate feature matrix as input for deep models, use Create_FeatureMatrix.py.
• Feature Vector V1: to generate the original feature vector for a trajectory to be used by GBDT model, use Create_FeatureVector_V1.py.
• Feature Vector V2: to generate the modified feature vector for a trajectory to be used by GBDT model, use Create_FeatureVector_V2.py.

• D-CRNN: this is our proposed model to perform driver prediction based on driving style information. This model combines several important compoenents of deep-neural-network architectures including recurrent, convolutional, and fully connected components. An implementation of this model in Tensorflow can be find here. Following diagram also describes the design of this model:

  

• VRAE: this is the Variational Recurrent Auto-encoder model, proposed by Fabius and Amersfoort (2015). Here we extend the original implementation to use it for driver prediction task by a modified loss function. The implementation of this model in Theano can be find here.

• GBDT: this is a Gradient Boosting Decision Tree model which we use it for driver prediction task. A Python imeplementatin of this method based on scikit-learn library is available here.

• CNN-model: this is a Convolutional Neural Network model to perform driver prediction task, which is proposed by Dong et al. (2016). Imeplementation of this method in Tensorflow can be find here.

• RNN-model: this is a Recurrent Neural Network model to perform driver prediction task, which is proposed by Dong et al. (2016). Imeplementation of this method in Tensorflow can be find here.

• Python: You may use either Python 2.7 or 3.
• Tensorflow: For all the deep models, except VRAE, you need Tensorflow (version >= 1.12.0 is recommended).
• Theano: To run VRAE model, you need Theano (version >= ??? is recommended).
• Cuda Library: You need Cuda Library for tensorflow and theano-based codes; version > 9.0.176 is recommended.

Note that our models can be run on both CPU and GPU machines.

Creating Feature Vector/Matrix

• Generate Feature Matrix: Use python Create_FeatureMatrix.py to create feature matrix. This will result in creating two files in data directory, one npy and one pkl file.
• Generate Feature Vector: This is to generate input for non-deep models, such as GBDT. For the original version of features as described here, use Create_FeatureVector_V1.py, and for the modified version as described in our paper, use Create_FeatureVector_V2.py. For each data file, we create one npy and one pkl file, using either of scripts.

Modeling and Prediction

• Run Deep Models: You just need to use python [MODEL_NAME].py. Make sure you have all the requirements satisfied.
• Run GBDT Models: Set the desired input data file in script and use python GBDT.py to run the model. Use --version v1 for the original feature vector (size 321), or --version v2 for the modified version (size 384).

You may find a raw sample file in data directory. In this file we have 5 drivers, and 10 random trajectories for each driver. The format of this file is described as follows:

Following is the result of different models on a random sample set of 50 drivers, with 50 trajectories for each driver. For deep models, we report accuracy on both segment as well as trajectory, see the paper for details.

• Sobhan Moosavi, Pravar D. Mahajan, Eric Fosler-Lussier, Colleen Saunders-Chukwu, and Rajiv Ramnath; "Driving Style Representation in Convolutional Recurrent Neural Network Models of Driver Identification", 2019