# Bicycle Dashcam Mark I


I love my bike - it is a workhorse I can park anywhere, a mid-1990s hybrid. After years of limited maintenance, in the past year, I've had to replace a tire, cassette, all the cables, pads, grips, and shifters. I've also just upgraded my headlight and taillight - the improvements that have been made in bicycle lighting over the last 15 years have been incredible.

I'm using my bike more this year - my downtown office recently moved to a building with badge access indoor bicycle parking and showers, with towel service, for cyclists - what a cool perk. So, I've been biking to work for the first time since I started at this company in 2006, 17 km down Yonge St in Toronto, about twice a week since June.

My rides have been great. Drivers along my route leave a lot of space. But it's hard to assess risk. The City of Toronto keeps detailed data on [cyclists killed or seriously injured](http://data.torontopolice.on.ca/pages/cyclists). There have been 11 KSIs on my route since 2008. But how do I compare that against, say, the risk of the 30 km drive to my Mississauga office? I've been rear-ended 3 times since 2010 commuting by car to Mississauga, but all have been at low speeds, only resulting in damage to my car - the consequences of getting hit on my bike are far more severe.

I was trying to think about what I could do beyond riding cautiously and ensuring I am visible. And, I have to say, a part of me is just always on the lookout for small, fun projects.

## Envision a bicycle dashcam

Bicycle dashcams have been done before, by [Cycliq](https://cycliq.com/) and others. But I envision something difference, a bicycle dashcam that could:

- Recognize the license plates of the cars around you. From a picture, it would look at the plates on all the cars, and then associate a plate number with the picture
- Record the speed of the cars around you
- Record the proximity of the cars around you
- A driver readable display, ie: "Driver ABCD1234, your current speed is 45". Like a mobile [Toronto Watch Your Speed](https://www.toronto.ca/services-payments/streets-parking-transportation/road-safety/vision-zero/safety-initiatives/initiatives/watch-your-speed-program/) program sign. _Would a driver allow a cyclist more space if they were aware their actions are being logged?_
- Log this data on a remote server
- Share this data, with a group. Perhaps associate "near miss" data from many cyclists, and identify troublesome areas, or troublesome cars.

## Introducing my Bicycle Dashcam, Mark I

![Bicycle Dashcam Mark 1](images/BicycleDashcamPrototypeRev1-1024x880.jpeg "Bicycle Dashcam Mark I")

My Mark I dashcam consists of a Raspberry Pi 3 with a Pi Camera (v1.3), a battery pack, running a small Node application which takes pictures, tries to recognize license plates with [OpenALPR](https://github.com/openalpr/openalpr), controlled through a phone friendly web interface.

![Bicycle Dashcam Dashboard](images/BicycleDashcamDashboard-1024x576.png "Bicycle Dashcam Dashboard")

## Testing and Results So Far

On the Pi 3, it takes between 8 and 800 ms to capture a photo with the Pi Camera, and another 7-8 seconds to run the OpenALPR license plate recognition process. I haven't looked into optimizing this, but I would be curious to see how fast this could get by adding a processor optimized for these tasks, like an [Intel Neural Compute Stick](https://software.intel.com/en-us/neural-compute-stick).

I've taken my prototype on a few drives, and a 5 minute bicycle ride. I don't know why I even tried using a Lego frame to mount the dashcam to my bike - it only held together for a few minutes of riding, and completely fell apart - I'll have to come up with something better for bicycle testing.

In the car, over a 30 minute drive (~120 photos) in traffic, about 15 license plates are identified. OpenALPR works exceptionally well - it can pick out the plate numbers even when it would be hard for a human to do so from the same photo. The limiting factor is the Pi Camera. At a stop, the pictures are fine, and OpenALPR will recognize the plates.

![Pi Camera image quality sufficient for OpenALPR when stopped](images/FineWhenStopped-300x225.jpg "Pi Camera image quality sufficient for OpenALPR when stopped")

However, as soon as the car is in motion, the image is washed out.

![Photo from Pi Dashcam while car is moving.](images/JustABlur-300x225.jpg "Just a blur. Photo from Pi Dashcam while car is moving.")

I have spent some time tweaking the photos taken by the Pi camera, trying out different modes. So far, I haven't been able to get great results.

As I look to take this further, I'll look at other Pi camera options, run further tests on my bicycle, perhaps move the project to a mobile phone app, as my phone's camera is significantly better than the Pi's. Also, I may explore inexpensive LED matrix screens for the driver readable display.

Source code: [https://github.com/raudette/plates](https://github.com/raudette/plates)