Architecture for a Self-driving Car’s “Vision”

Development of an application architecture for automated image recognition for a self-driving car using sensors and machine learning

Client situation:

Self-driving cars need to “see” the road as you would and need to recognize and adapt to new situations in real time.  Our client was using a manual validation process to “teach” their self-driving car how to recognize its environment, a lengthy and potentially error-prone approach.  They engaged Saguaro to develop a more sophisticated and automated system to increase efficiency and accuracy.

 What Saguaro did:

The client’s self-driving car was equipped with cameras and a RADAR/LIDAR system that detected various objects in the road, producing > 11 million images to analyze and correct.  Our engineers developed the architecture of a complex distributed solution including a web-based tool for manual and semi-automated ground-truth data validation. This web-based tool included a server application, a pre-processing unit, and a graphical user interface (GUI) that would provide the main interface for human testers. Our engineers developed a solution that could automate parsing of the recordings into individual images and then applied software tools that presented and edited the images very rapidly and accurately. In addition, Saguaro provided data inspection for, and delivery of, validated ground-truth data.

Client benefit:

With Saguaro’s input, the client was able to automate an otherwise lengthy and potentially inaccurate process, improving productivity and reducing time-to-market for their product.

Languages & tools used in this project: HADOOP clusters; state-of-the-art JavaScript frameworks for the user interface; 3 different backend approaches based on Windows Server 2012 R2 using C#, Linux using Java, and Linux using Python;Log4Net; XML; MD5 encryption for DB credential; HTML5; CSS

 

 

Back to case studies