A Tire-Road Friction Coefficient estimation method based on camera-informing and dynamic analysis.

This part gives a classification result of the surface by analyzing the texture of the surface through GLCM((Gray Level Co-occurrence Matrix)) and extracting fueature by pretrained ResNet. Later we will use this classification to inform the dynamic model.

• /data: ground truth videos, classified images(training, dev and test data);
• /data/dataset.py: Pytorch Dataset Class, used for Dataloader;
• /data/data_preprocess.py: Convert videos to imgs and split into train, dev and test sets;
• /model: trained models and checkpoints;
• train.py: main training script;
• evaluate.py: run model on test set to evaluate it;
• model.py: definition of the NN and the layers;
• util.py: helper functions, like the implementation of glcm.

To train the model, run (set up the WANDB_API_KEY in util.py if using wandb):

python3 train.py --use_wandb --experiment_name="your-experiment-name" --batch_size=40 --lr=0.001 --lr_decay=0.1 --max_epoch=40 --num_workers=12 --drop_out=0.1

To evaluate on test, run:

python3 evaluate.py

• 06.23 Using ResNet50

• 06.30 Using ResNet18

• 06.30 Using TensorScript for Acceleration

• 07.02 Stream Clustering using DBStream