Non-Conventional Traffic Participants for Semi-Autonomous Vehicles

Подробнее о видео

Presentation for IEEE ANTS 2020 Workshop on AI/ML for Connected Vehicles

Abstract: In recent time there is plethora of progress on both research and commercialization aspects of autonomous vehicles, mostly for European and North American highways. However, traffic participants at developing countries and specific use cases like in an airport require a machine learning system to be trained with unusual situations. For example, responding to presence of animals in a sub-urban road, or unusual shaped vehicles at an airport often poses challenge due to lack of enough training data, privacy and security aspects of the environment. This talk will present results on comparing existing Convolutional Neural Network (CNN) models with respect to latency and accuracy for unusual traffic participants in Indian road data set. We shall propose use of VR digital twins to prepare synthetic datasets. In the context of training CNN models with real and synthetic data sets, we shall investigate how a CNN actually works and compare the working of intermediate layers of CNN using data visualization techniques. Finally, we shall look at the human machine interaction and teaming aspects of semi-autonomous vehicles and present case studies on operating a vehicle using multiple modalities.