Development of Plausiblity methodologies for Deep Neural Networks: Towards Functional Conformance

• Plausiblity Verification for 3D Object Detectors using Energy based Optimization: Framework to reduce False Positives using cross sensor validation.
• Knowledge Integration Plausible Motion Forecasting: Predicting conformal trajectories with kinematic and environmental priors.

Advancing functional safety for deploying Neural Networks in Autonomous Driving.

• Created a motion prediction model that forecasts trajectories of dynamic actors around the ego vehicle, leveraging fused LIDAR point clouds and HD maps through a novel sensor fusion architecture.
• Pioneered a multi-modal cutin detector using a CNN and Attention-based model to anticipate lane merges of surrounding vehicles by fusing RADAR and Camera data.

Model Uncertainty estimation on Semantic Segmentation Network with a real time deployment on Nvidia Drive PX2 for Autonomous Vehicles.

• Parallelization of MC Dropouts for real time sampling of Uncertainty.
• Custom layer optimization of Network model using TensorRT optimizer.
• Deployment of optimized model into Drive PX2 for real time segmentation.

Worked on,

• Design of Segmentation Network architectures for Deep Neural Network.
• Development of training Pipeline for DNN using TensorFlow.
• Uncertainty Estimation of DNN models resulting in a boost in performance of the pixel wise classification accuracy.
• Benchmarking of IBM cluster GPUs for training performance evaluation.