ANSYS SCADE Vision Capabilities
Autonomous Vehicle Perception Software Robustness Testing
SCADE Vision powered by Hologram automates the identification of potential vulnerabilities in autonomous vehicle (AV) perception systems. It helps you to find and categorize edge cases where these systems might exhibit unsafe behavior. It is intended for safety analysts, neural network developers and anyone else who works with machine learning-based perception systems.
SCADE Vision enables automated testing of the AI-based AV perception software under test (SUT), usually a convolutional neural network (CNN). Testing consists of running the SUT inference algorithm twice against each raw input video captured from the AV sensors: the first inference is run on the baseline, unmodified frames, while the second inference is run on an augmented/modified version of the input video frames, when there are objects of interest (e.g., pedestrians, cars) detected in the scene. The SCADE Vision engine then analyzes the SUT outputs stored in the results database using several defect analysis algorithms to identify weaknesses and fragilities in the AV perception software, including weak detections or false negatives.
SCADE Vision does not require labeled data to support AV perception software testing; instead, it searches through raw sensor data recorded by the autonomous vehicles.
A web-based UI provides you with the ability to configure (input video/images repository, perception algorithm/CNN under test, etc.) and launch one or several analyses. The UI also helps you to quickly identify the input videos and frames which highlight probable defects of the AI-based AV perception software, making it very effective to browse through the AV data lake.
Triggering events identification in autonomous vehicle perception software
Once the SCADE Vision engine has performed the analysis of the AV data lake, the web-based UI supports analysts in categorizing the probable defects of the AV perception software into proposed triggering events, or root causes, for fragilities in the AV perception systems.
Triggering events are diverse: weather conditions (snow, rain, wildfire), lighting conditions (glare, night, high beams), infrastructure (fences, reflective surfaces, statues), types of road users (wheelchairs, people in costumes), or simply an incomplete training of the AV machine learning systems.
Automatic safety report generation
SCADE Vision enables you to automatically produce web and printable reports for perception algorithms development and safety teams. A dedicated report generation UI helps analysts to provide commentaries, including ideas for mitigations, on key triggering events along with example defects.
Automatically generated reports help you to structure and communicate the results of the safety analysis with members of the AV perception software development team and other interested parties, in a virtual feedback loop.