Part 1: Rally cars and AI

February, 2019

Author: Marcus, Furqan

Objective: Automate indexing of videos from car races.

Duration: 2 weeks

Together with Nordur Film (https://www.nordurfilm.no/) we are exploring the area of making video production from race car races a blast. Each race has about 80-150 participants. Imagine that you 

Potential ways to solve the challenge.

OCR

OCR (Optical character recognition) is widely used for detecting text from written articles, road side signs images etc. Our first step towards detecting the number on the cars was through the OCR services provided by Microsoft, Amazon and Google. Additionally we tested a few other common algorithms for this job. Unfortunately none of them could solve our challenge.

ANPR

ANPR (Automatic Number Plate Recognition) is widely used in many services such as parking and perimeter security. Through a couple tests with one of the most common solutions we could see a prominent potential. There is only one challenge: Race cars don’t commonly have license plates..

Deep learning

After exploring existing solutions we concluded that deep learning was the way to go. Our challenge now is to gather test data.

JTO: The Joy Tech O...

Our software for sampling videos of people from races has now been modified to also generate test data for our deep learning algorithm. The tool was developed in order to increase efficiency and minimize the probability of human errors.

Blur

Despite filming in 50 frames per second most of the frames will still have problems with blur in the images. Blur is known to be one of the larger challenges when it computer vision.

 

4 learnings from Rally Sørlandet

1. Existing ANPR only solves part of the problem.

Using only ANPR will only give results on approximately 60% of all rally videos. Additionally, it will not work for most other motor sports.

2. Blur is a serious challenge

As most car numbers are on the side door or side window and most cameramen are standing right next to the road there is only a split second where the number can be captured. During this split second the numbers are often blurred out.

3. Sorting videos from car races has a big potential

Producing personal videos from car races creates a lot of value for the drivers, fans, sponsors and the sport in general.