This is an adapted variaiton of RNN-RBM for unsupervised vehicle trajectory pattern learning on a multi-lanes highway. The dataset used is US Highway 101.
Unlike the original RNN-RBM, this is a model which performs well for real-valued dataset. After training, it can memorize a given vehicle trajectories pattern on a specific route, which can be used to reconstruct vehicle trajectories based on a input dataset polluted by noise.
The structure of RNN-RBM(Nicolas Boulanger-Lewandowski et al.) is illustred below:
Put the input trajectories dataset (eg. "trajectories-0750am-0805am.csv" of US Highway 101) in the directory "./ngsim_data/ "
-
python ngsim_manipulation.py# pretreatement of dataset -
python weight_initialization.py# pre-training of RBM Neural networks -
python rnn_rbm_train.py <num_epoch># training, num_epoch can be about 10 -
python rnn_rbm_reconstruction.py ./parameter_checkpoints/<fichier .ckpt># reconstruction of trajectories
The output picture is stocked in "./picture_folder".
There are different versions of ngsim_manipulation, as well as for rnn_rbm, RBM, draw, they deal with different situations. For using them, rename them(delete the version notation).
One demonstration of noise reduction:
Further work can be conducted, for example, changing the standard RNN by LSTM, or changing RBM by DBN, etc.