Semi-supervision with Centerness in 3D Point-cloud object detection for Autonomous Vehicles

Autonomous driving has been a very active area of research in the current eraand many different proposals is making it’s way in the research field to makethe transition from current driving assisting models to an autonomous drivingone. Recent works in perception for AV emphases on the low accuracy issues indeployment of AV in industries. Lower perception capability of vehicle indicatelower effectiveness of the decision making module which in turn leads to lowerconfidence in executing self-driving scenarios. A high amount of research hasfocused on LiDAR point cloud detection assuming the LiDAR provide accurateground-truth while some propose using pseudo LiDAR projections from the imageby predicting depth over an image. Yet the detection models have not reached theindustry requirement of per-cent accuracy. Our project aims at introducing theconcept of centerness to PointRCNN network to achieve higher order of accuracyjust using LiDAR Point-clouds. Preliminary training provide insights on higherpotential of this method and promising results over state of the art methods.

• Improved the perception system of autonomous vehicle with average precision and average recall (IoU=0.7) for vehicle class exceeding the baseline PointRCNN model by 20~25% (on argoverse dataset) and achieved considerable improvement on KITTI benchmark dataset as well with AP of 92.02 for car class.