Decoupled 3-D object detector

This paper proposes an efficient cascaded 3-D object detection architecture. Such an architecture decouples the 3-D object detection pipeline to maximise the utilisation of the inherent advantages of RGB images and LiDAR point clouds in order to perform 3-D object detection while maintaining low computational complexity. Our proposed architecture relies on a cascade of two networks, the first leverages the texture density in images and the maturity of state-of-the-art 2-D object detectors to classify and obtain initial region proposals for objects in the scene. These proposals are fed to a lightweight secondary network that leverages the compactness of bird-eye view point cloud representations to perform orientation and size estimation. The 3-D bounding box proposal is constructed by fusing predictions inferred from both networks, as predictions lie on orthogonal planes. Evaluated on the KITTI benchmark data set, we show that the proposed method obtains results on-par with more complex end-to-end 3-D detection methods while greatly reducing computational and memory requirements. This work also presents results from the deployment within a perception pipeline, and analyses challenges faced in deployment within a frontal perception pipeline. © 2022 Inderscience Enterprises Ltd.