NOTE FOR JUDGES:

If our video presentation can be more than 3 minutes (but less than 4 minutes), please watch our Extended Version of the presentation submitted in the "Try It Out" links instead of the one in the official "Video Demo Link" spot above.

What it does

Our app will watch videos (that you provide) of traffic and output the mileage and predicted annual fuel costs of each car. This data is useful for strategically targeting environmental campaigns (commercial or governmental) toward the population that would be most likely to make positive changes. For example, those that pay more in gas each year would probably be more likely to switch to an electric vehicle. Monitoring this data over time can yield even more insights, such as the progression towards electrification in different parts of the US.

How it works

We built a deep convolutional neural network model in TensorFlow to identify cropped images of cars. We trained the model with the existing Stanford Cars Dataset[1] which has over 16,000 images of 192 different types of cars. Additionally we are using a preexisting, You only look once (YOLO)[2], object detection machine learning model to identify and draw boundary boxes around specific cars in video. The ML model outputs the types of cars in the video, then our Node backend queries the FuelEconomy.gov api to determine the estimated mileage of each car, and then an algorithm is run to find the effective mileage based on the speed each car is traveling (since mileage varies depending on speed). All of this data is put together and predicted annual fuel costs are calculated based on that mileage and the average number of miles driven per person, which is 11,500. This data is sent to the frontend which very nicely displays it to the user.

Challenges we ran into

We struggled to get the ML model working to identify the cars. Even after getting it working, its accuracy seems to be very low. With more time, surely we would be able to significantly improve it. Previous literature had show good results with a small amounts of car classes[3].

Accomplishments that we're proud of

Just the fact that we finished this project and were able to make a strong presentation makes us very proud, especially given that this was only 24 hours and many of us had other commitments during the day.

What's next for Automotive Mileage Monitor

We'll need to expand the car dataset to be more representative of the whole US, since it's skewed towards higher cost vehicles right now. Then, theoretically, the rest of the project (relying on the EPA's FuelEconomy.gov) should scale automatically.

References

[1] J. Krause, M. Stark, J. Deng, and L. Fei-Fei, “3D object representations for fine-grained categorization,” 2013 IEEE International Conference on Computer Vision Workshops, 2013.

[2] J. Redmon, S. Divvala, R. Girshick, and A. Farhadi, “You only look once: Unified, real-time object detection,” 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016.

[3] Y. O. Adu-Gyamfi, S. K. Asare, A. Sharma, and T. Titus, “Automated vehicle recognition with deep convolutional neural networks,” Transportation Research Record: Journal of the Transportation Research Board, vol. 2645, no. 1, pp. 113–122, 2017.

Share this project:

Updates