Robust Perception is vital in automotive Cyber-Physical Systems (CPS). Although the supporting technologies have advanced recently, enabling robust perception remains challenging for researchers and industry alike. The highly variable scenarios in complex urban environments can lead to erroneous perceptions, which are factors in most driver-related crashes. In this chapter, we present our experience developing AV perception models capable of better understanding driving scenes, thus improving their robustness. Specifically, we propose using scene-graphs as a better Intermediate Representation (IR) for road scenes. Besides, we develop a novel spatio-temporal graph learning approach based on scene-graph representations for modeling the risk of driving maneuvers. Our approach better understands driving scenes and converts them into an estimated risk level by leveraging a network architecture consisting of a Multi-Relation Graph Convolution Network (MR-GCN), a Long-Short Term Memory Network (LSTM), and self-attention layers. We demonstrate how a scene-graph approach for AV perception enables the AV to better assess risk across various driving maneuvers than state of the art, thus being more robust. Moreover, our approach can more effectively transfer knowledge learned from simulated data to real-world driving scenarios. Lastly, we show how adding spatial and temporal attention layers to our approach improves its explainability.
Scene-Graph Embedding for Robust Autonomous Vehicle Perception
Machine Learning and Optimization Techniques for Automotive Cyber-Physical Systems ; Chapter : 18 ; 525-544
2023-03-27
20 pages
Article/Chapter (Book)
Electronic Resource
English
METHODS FOR SPATIO-TEMPORAL SCENE-GRAPH EMBEDDING FOR AUTONOMOUS VEHICLE APPLICATIONS
European Patent Office | 2023
|METHOD FOR GENERATING A WIDE-AREA PERCEPTION SCENE GRAPH
European Patent Office | 2018
|