Video based car Radar

Video based vehicle radar built entirely from scratch in the C# programming language.
1. A database consisting of 2000 car/non-car images were trained using a genetic algorithm that was wrapped inside the ADABoost meta algorithm. 150 pictures were manually created from photos (still video frames), whereas the rest were automatically generated from previous attempts and classified into positive/negative by inspecting them with a human eye.
2. 60 features each consisting of one of 8 simple rectangle types were constructed during approximately one hour of CPU time on a Core i7 based computer.
3. A Gopro Hero 2 camera filmed the Swedish highway (E4) in 1280x720 (720p) resolution
4. Each pixel within a certain region of the each picture frame scanned about 10 sub-windows of different sizes (viewport distance dependent). The 60 features formed a linear combination "vote" and marked a rectangle if a car was identified.
5. Identified rectangles were post-processed to form clusters of different observed vehicles.
6. The identified clusters were recursively intersecting other "spanning clusters" within 8 previous video frames, if the cluster search reported more than a certain amount of "cluster activity" going on the identification finally was over and a vehicle was detected.
7. The radar lines just above the cars bonnet show distance to vehicles obeserved. Blue lines to the right is meant to be used for tracking of distance (following at a set distance "adaptive cruise control". left white lines are for informational purpose only. System looks at the own cars viewport to determine whether a vehicle should be classified as "white" or "blue".

Implementation time about 200 hours, spent approx 150h for the machine learning part and about 50h for striding and post-processing the video.

So, one might ask why do you make your own vehicle radar in your spare time?
BECAUSE I CAN!

• Category

  Science & Technology

[Ben Wenger]

Have you tested the robustness of your approach. These are ideal conditions, but how does it perform under hard shadows like sunrise/sunset? Or rain/fog? What about sun glare causing the aperture to shrink and white balance to adjust? Just something to think about. These become most important after the ideal conditions are shown to work (which you have shown nicely). Also, you may want to do some analysis of false positives/false negatives/true positives/true negatives if you want to publish.

[Gustaf Sundberg]

I think its a matter of collecting sample pictures from less ideal conditions.

Let it run using the current detector and then retrain the system again.

Each of the single features I have trained has a metric object with Precision, Recall, error, weighted error and more.

Recall = What percentage of positives are being recognized

Precision = What is the accuracy of the ones that where detected as positive.

The hardest few percentage is being handled in the post processing step. (3d analysis)

[Gustaf Sundberg]

(more text)

When I combine 20-30 or more single features, the combined Precision and Recall goes very high. 100% on the training set.

[scorpio103066]

This should be on Kickstarter - count on my pledge on day 1.

[Gustaf Sundberg]

What is Kickstarter?

[TedRobotBuilder]

German Autobahn?

[Gustaf Sundberg]

Sweden.

[NEETchannel]

any chance of this being open source?

[Gustaf Sundberg]

Iam aware that strictly speaking RADAR stands for "Radio Detection and Ranging".

Since Iam entirely relying on a Video camera (GOPRO) and not Radio or GPS, Lasers or anything else I guess its a VIDAR then.

However most people are familiar with the concept of what a RADAR does, this system does exactly detection (of vehicles) and gives you a range to them. Hence the prefix "Video based".

[Ralphie Dee]

That is awesome... Nice build!

[94DarthNihilus]

Wow, that is amazing. Is there a paper or a web-site where you describe what you did?

[Gustaf Sundberg]

(more text)..

The actual "knowlegde" of the trained classifier is represented by 60 * 32 bits. (240 bytes)

Its a very compact representation of being able to "see" vehicles, but remember, the actual search space is staggering: 2^1920 or about 9.5e+577 combinations!

Because of the vast amount of combinations, a Genetic Algorithm was built to train the system.