Autonomy: Pedestrian Detection
Synthetic data for computer vision to enable more capable and ethical AI.
Robust pedestrian monitoring is critical for ensuring the safety and performance of driver assistance systems and autonomous vehicles. Predictive pedestrian detection is a form of automotive AI that can help computer vision systems interpret, prepare for and respond to any number of scenarios, with or without driver oversight.
Obtaining training data is difficult given safety risks and the potential for large numbers of edge cases. The most obvious example is a child running in front of a vehicle, a scenario for which it is unethical and impractical to create the necessary training data. Synthetic data enables computer vision engineers to create millions of simulations of such situations, informing the development of critical safety systems.
Pose Estimation
Complete control over any possible human pose, all with accurate body landmarks to help developers build more robust pose estimation models for pedestrian monitoring. A large library of actions and movements is mapped to thousands of identities and body types.
Multiple Pedestrians
With Synthesis Scenarios, computer vision engineers can generate labeled images and video datasets for complex scenes with numerous pedestrians. The system provides flexibility to plan for and simulate edge cases and rare events.
Confounds
ML models need to account for pedestrians who are partially or fully occluded by objects such as bicycles, other vehicles, elements in the environment, other people, and much more. Synthesis Scenarios can generate data for situations when one or more people aren’t 100% visible to a vehicle’s onboard CV system.
Attention Tracking & Gaze Estimation
Pixel-perfect annotation of gaze angle can help develop more accurate models. Gaze estimation is critical for assessing a pedestrian’s attention and determining whether they’re looking at either the vehicle or a vehicle occupant, including the driver.