This is a school project done with a company to classify dangerous vs. non dangerous driving on the road. I strted with some exploaration and preprocessing and moved to base models ( Random forest and gradient boosting ) and then tried resnet based on this paper+repo:
https://github.com/hfawaz/dl-4-tsc
THe data used for this project belong to the company I did the project for, hence could not upload it here. It contains the following:
- Accelerometer(Csv file @ 50Hz)
- Gyroscope (Csv file @ 50Hz)
- Magnetometer(Csv file @ 50Hz)
- GPS (Csv file @ 1Hz)
THe following files contained modeling and exploratory analysis:
- Analyzing GPS Data.ipynb: an interactive map of GPS data to see any patterns based on location
- modeling accel with random forest_-animation of video with bus-.ipynb: modeling with randomforest using only accelerometer
- modeling accel with gradientboosting_-animation of video with bus.ipynb: modeling with gradientboosting using only accelerometer I created frames for a gif showing the gradientboosting predictions https://imgflip.com/gif/33dp6h
- modeling accel and gyro with gradientboosting.ipynb: modeling with 2 sensors
- modeling GPS accel and gyro with random forest.ipynb: modeling with 3 sensors
For the base models all python packages needed are listed in requirements.txt file and can be installed simply using the pip command. this project was running on Google Deep Learning VM Based on: Debian GNU/Linux 9.8 (stretch) (GNU/Linux 4.9.0-8-amd64 x86_64\n)
- numpy
- pandas
- sklearn
- scipy
- matplotlib
| Gradientboosting | Random Forest | |
|---|---|---|
| Accelerometer | acc: 79.69% recall: 84% precision:80% | acc: 82% recall: 89% precision:81% |
| Accel+Gyro | acc: 79.32% recall: 85% precision:78% | acc: 81% recall: 87% precision:79% |
| Accel+ Gyro +GPS | acc: 82% recall: 88% precision:81% | acc: 81.57% recall: 88% precision:79% |