3D Object Detection in Point Cloud Using Key Point Detection Network

This paper improves an existing model which is used in real time to detect an object in the 3D point cloud. We have used ResNet-based Key point Feature Pyramid Network (ResNet-KFPN) which predicts 8 perspective key points of a 3D bounding box in the image space which is different from other models, and then we use geometric relationships of 3D and 2D perspective to restore size, location, and position in 3D space. We have used the image in the form of Bird’s Eye View (BEV) as input to the model which embeds the intensity, height, and density of the particular point of the point cloud. Firstly, we convert LiDAR points into BEV in the form of an image, and then apply standard Convolutional Neural Network (CNN) for object detection. We have also used multiple loss functions with adaptive learning rate so that the model can learn more accurately.