Camera-LIDAR Integration: Probabilistic Sensor Fusion for Semantic Mapping

An automated vehicle operating in an urban environment must be able to perceive and recognise objects obstacles a three-dimensional world for navigation path planning. In order plan execute accurate sophisticated driving maneuvers, high-level contextual understanding of the surroundings is essential. Due recent progress image processing, it now possible obtain high definition semantic information 2D from monocular cameras, though cameras cannot reliably provide accuracy 3D provided by lasers. The fusion these two sensor modalities can overcome shortcomings each individual sensor, there are number important challenges that need addressed probabilistic manner. this paper we address common, yet challenging, LIDAR/camera/semantic problems which seldom approached wholly Our approach capable using multi-sensor platform build voxelized map considers uncertainty all processes involved. We present pipeline incorporates readings (cameras, LIDAR, IMU wheel encoders), compensation motion vehicle, heuristic label probabilities images depicted Fig. 1 . also novel efficient viewpoint validation algorithm check occlusions within camera frame. A projection performed LIDAR point cloud. Each labelled scan then feeds into octree map-building updates class voxels every time new observation available. validate our set qualitative quantitative experiments USyd Campus Dataset. ref-type="fn" rid="fn1" xmlns:xlink="http://www.w3.org/1999/xlink">1 These tests demonstrate usefulness evaluating performance perception system typical autonomous application.