Lidar 3D Object Detection

State-of-the-art 3D object detection on lidar data pretty much requires applying a deep learning model. Most models are trained in a supervised manner with lidar scans as inputs and 3D objects as outputs. I would say there isn’t anything too special about these models, except in how they handle point clouds which are unordered set of points.

I wrote a ROS 2 package that uses MMDetection3D API for lidar-based 3D object detection. The API supports a variety of lidar object detection models, which can be swapped around. This is the documentation about the same software which has I have integrated into Nova’s perception stack.

Here are some demos on the KITTI 3D object detection dataset viewed through RViz:

A video demonstrating a self-driving car from Nova using this software to detect a bicycle around a parking lot at UTD is available below: