Driveablespaces Estimation Group(Computer Vision)
​

• Implementing and validating algorithms such as”Object-Contextual Representations for Semantic Segmentation” and ”EfficientPS

• OCR + HRNET backbone for Full Segmentation of Images on Berkley Deep Drive Dataset

• Building data loading pipeline for inference on the robot.

• Achieved Full Image semantic segmentation with inference speeds about 4 Images/sec.(1280 x 720 on V100 GPU)(Pytorch before any optimization)

• Increasing Inference speed using TensorRT, quantization and model pruning.

Tools Used- Pytorch,AWS,numpy,Python.

​

 

Localization Group

​

• Implementing Mapping and localization using LEGO-LOAM geometric SLAM.

• Implementing localization sub-module using LOAM feature detectors.

• Implementing GPS-based localization for the Bot.

• Optimizing map loading for the robot by loading only required subspace in maps.

• Decreased memory requirement for maps.

Tools Used- ROS,PCL,C++