Research Projects

Humans driving vehicles often rely on the traffic scene visuals, locality knowledge and acquired cognitive skills. Inspired by the same, we propose deep learning model that can accomplish autonomous driving by accessing camera visuals, GPS and inertial measurements. The trained model can make path prediction in real-time while avoiding enormous amount of computing and high speed communication. This should facilitate large scale adoption of autonomous vehicles without burdening smart vehicle electronics, transportation infrastructure and communication network, while holding desired functional accuracy and performance. PIs propose to encode the traffic context data and use the encodings to obtain conditional distribution via a variational autoencoder network. The trained network will therefore be able to predict the self-driving vehicle movements given the traffic context. The model will be optimized to operate on Indian traffic scenarios.