Nathan Kurtz, Autoliv
In the world of automated driving, sensing accuracy is of the utmost importance, and proving that your sensors can do the job is serious business. This is where ground truth labeling has an important role in Autoliv’s validation process. Currently, annotating ground truth data is a tedious manual effort, involving finding the important events of interest and using the human eye to determine objects from LiDAR point-cloud images. This talk presents a tool that was developed in MATLAB® to alleviate some of the pains associated with labeling point-cloud data from a LiDAR sensor and the advantages that tool provides to the labeler. The capabilities of the tool are discussed, including assisting users in visualizing, navigating, and annotating objects in point-cloud data; tracking these objects through time over multiple frames; and then using the labeled data for developing machine learning based classifiers. The talk also describes how the output of the labeling process is used to train deep neural nets to provide a fully automated way to produce vehicle objects of interest which can be used to find false-negative events. To do this with a human analyst takes as much time as to play back the entire data set. However, with a fully automated approach, it can be run on many computers to reduce the analysis time. This video shows the time savings as well as the accuracy of the labels achieved and how this approach provides substantial benefit to Autoliv’s validation process.
Recorded: 2 May 2018
You can also select a web site from the following list:
Select the China site (in Chinese or English) for best site performance. Other MathWorks country sites are not optimized for visits from your location.