Empowered by a comprehensive data set of U.S. bridges and tunnels, RV Nav allows users to safely reach their destination without opening up their vehicle like a tin can. Plan your trip ahead of time with our turn-by-turn directions. We’ll also try to notify you of potholes along the way!
Language: Python(pandas) Frameworks: Fast.ai Services: AWS
Intro and Onboarding
RV Nav was an 8-week endeavor into the product development cycle, allotting a team of developers a window of time to onboard an agile but flexible product strategy while meaningfully contributing to an unseen codebase. For me, this meant immersing myself in a new mindset and being willing to openly collaborate with a talented team of web, UX, and data science developers.
The data science team took on three separate tasks: increase the number of bridges in the database, prototype an image classification neural network to identify road hazards, and build an in-house pipeline that provided RV related insights from reviews on the web.
The Data
Bridges and Tunnels
Our sole source was decided to be The U.S. Department of Transportation: Federal Highway Administraion. They have a robust and current public database for both bridges and tunnels. We restricted the clearance window from a lower bound at 7 feet to an upper bound of 20 feet, reformatted the GPS coordinates to a decimal format to be compatible with ARCGIS, and dropped the remaining unused columns. In total, roughly 80,000 new bridges were added.
Roadway Images
Images of varying roadway content were gathered from the web including pictures of potholes, animals, oil spills, traffic, and unoccupied streets.
Buisness Reviews
Roughly 500 items of text related to RV’s and travel were collected from popular app distribution sites.
Backend API
Using Amazons Elastic Beanstalk to host our API, we redeployed the updated source package to include the new bridge data.
Classification Neural Net
Following along with tutorial videos and notebooks from Fast.ai, we were able to quickly produce a trained convolutional neural net that boasted a ~95% accuracy score on the test data for images containing road hazards. Additional layers were trained on top of the resnet18 model and validated against a data set of 200 images unseen in training.
NLP pipeline
Sentiment analysis was performed and visualized on the review data.
(Scattertext by teammate George Hou) Link to interactive.
Retrospective and TO-DO’s
I’m glad to have had the opportunity to work in a production environment. Almost every deliverable we produced here I would like to take to the next level. For the neural net, it would be cool to make the jump to object detection in near real-time from mounted hardware and apply more extensive testing and evaluation. Building an automated bot to crawl troves of street image cataloging bridge heights would be a fun challenge. Deploying a generalized tool for text analysis and visualization is something I’d like to work on as well.
Here’s a rv-navi-GATOR for posterity
