基于深度神经网络的异常驾驶检测内含数据集-含说明书(可运行).zip

# Abnormal Driving Detection Based on DNN [](LICENSE) [](https://github.com/Lmy0217/AbnormalDrivingDetection/pulls) [简体中文](README_zh.md) This implements training and testing of some models from [Abnormal Driving Detection Based on DNN](paper.pdf) by *Mingyuan Luo* and *Xi Liu*. ## Environment - Operating system: CentOS 7.2 - Data would take up to 260GB disk memory - Memory cost would be around 40GB - Dependencies: - [CUDA](https://developer.nvidia.com/cuda-toolkit) and [cuDNN](https://developer.nvidia.com/cudnn) with GPU - [Torch](https://github.com/torch/torch7) with packages ([nn](https://github.com/torch/nn), [cunn](https://github.com/torch/cunn), [cutorch](https://github.com/torch/cutorch), [cudnn](https://github.com/soumith/cudnn.torch)) installed by default, as well as some special packages such as [image](https://github.com/torch/image) If you already have Torch installed, update `nn`, `cunn`, and `cudnn`. ## Prerequisites - Download Abnormal Driving Detection ```bash git clone https://github.com/Lmy0217/AbnormalDrivingDetection.git cd AbnormalDrivingDetection ``` - Download the [State Farm Distracted Driver Detection](https://www.kaggle.com/c/state-farm-distracted-driver-detection/data) dataset and extract the zip file (imgs.zip) in the folder `datasets` (now, this folder should contain two folder named 'train' and 'test' respectively) ## Training The training scripts come with several options, which can be listed with the `--help` flag. ```bash th train.lua --help ``` To run the training, simply run train.lua. By default, the script runs vgg16 on dataset with 10 epochs. To train wide12 with 20 epochs: ```bash th train.lua -model wide12 -nEpochs 20 -batchSize 4 -LR 1e-2 -weightDecay 1e-4 ``` There are a variety of the options of model, please see the folder [`models`](models). Every epoch trained model will be saved in the folder `results/[model]` and all losses saved in `results/[model].log`. ## Testing The testing scripts come with several options, which can be listed with the `--help` flag. ```bash th test.lua --help ``` By default, the script tests the 0 epoch of vgg16. To test the 11~20 epochs of wide12: ```bash th test.lua -model wide12 -index 11 -more 10 -batchSize 8 ``` Every tested result will be saved in `results/[model]_test/[model]_[epoch]_test.log`. Every log file including two lists for `k` and `Top-k`. ## Some results The trained models with learning rate 1e-2 achieve better accuracy rates: | Model | Top-1 accuracy | Top-5 accuracy | | -------------- | -------------- | -------------- | | VGG-16 | 99.3223 | 99.9197 | | Wide-12 | 99.3580 | 99.9643 | | Cardinality-25 | 99.2153 | 99.9554 | | ResNet-16 | 99.4560 | 99.9910 | | DenseNet-88 | 99.3669 | 99.9465 | | DWC-88 | 99.2688 | 99.9375 | | DWCR-88** | 9.8528 | 50.6286 | | DWC2R-29 | 99.2331 | 99.9286 | ** DWCR-88 is failed. The trained models with learning rate 1e-4 in 10 epochs achieve accuracy rates (dotted lines are losses and full lines are Top-1 accuracy rates):  The more details of results of trained models (learning rate = 1e-4) are saved in the folder [`ours`](ours). ## License The code is licensed with the [MIT](LICENSE) license.