Car system for external object detection

20 Thousand Leagues was tasked with building an object detection system as part of Advanced Driver-Assistance System. The solution is aimed at cars which do not have an object detection system of their own, as it would reduce the risks associated with the use of otherwise good vehicles. 

Challenge

Object detection needs to differentiate between possible dangers, such as cars, trams, pedestrians, cyclists. To warn about a possible accident, the should recognize signs, road markings, and traffic lights. A bad read on them often causes crashes, so the system needs to tell where those choke points are.

As an external product, the object detection system has to be financially worth implementing. This requires software that is not demanding on the computing power of cameras installed in cars.

Object detection should tell cars from motorcycles, people from animals, road signs from buildings. All with light software.

Machine learning taught the system how to differentiate between objects. Computer vision is employed to spot them and warn the driver.

Car-sharing companies may run their cars for longer. Freight transportation is more secure.

Solution

Our system employs computer vision to classify objects based on their features, much like humans do. Head, body, and four limbs with only two designed for walking make a human. Multiple doors, side mirrors, a registration plate, and least one human in this moving box make a car. Motorcycles are not that different, and the lack of side mirrors would indicate we are dealing with a bicycle.

Machine learning methods were used to make the system recognize what is what and who is who. The algorithm behind the system went through thousands of photos, and constant adjustments from engineers made it precise enough for real-life use. Unlike humans, it can’t be distracted and it does not get tired.

The system is built with low calculating power in mind. It is based on Python, and Mask R-CNN was the model for machine training. The load is negligible because while assessing the situation in real-time, our solution doesn’t have to handle more than one image at a time.

Building software is a team sport. Without proper coordination, communication and verification; you’ll end up with a sub-par product and a whole bunch of people dissatisfied. This is also why we at 20kL are frank, upfront, transparent and direct with our customers.
R9707ci8TTSLAfl0Ct8mug

Simon Kaastrup-Olsen
Chief Executive Officer

Results

As the solution was ordered by an IT company rather than a car manufacturer, it has been further redistributed to other businesses. This includes a large car-sharing company that uses the object detection system to lower the number of client-caused accidents and avoid downtime on cars involved in those accidents.

The system can also be used in freight transport. An operator with our solution installed in their trucks has had an increase in clients looking to outsource their logistics.

Technology: Computer Vision, Machine Learning
Tools and framework: Python, Mask R-CNN

We want to build the best software we can, we do that by working with our customers by asking lots of questions, visualising every possible bit, transparent coordination and no delayed decisions.
R9707ci8TTSLAfl0Ct8mug

Simon Kaastrup-Olsen
Chief Executive Officer

Next story

Repair Automation

Rope Robotics