Continuous Segmentation of Pointcloud

Autonomous vehicles is an actively developing technology which requires advanced solutions that will allow real-time vehicle environment analysis. Using a set of sensors, vehicle receives a large amount of information about the environment, which allows to predict the further movement of road users and plan next actions. One of the most used sensors in the industry today is lidar (LiDAR), a sensor that quickly rotates and illuminates surrounding objects with a laser, building a sparse three-dimensional representation of the surrounding world, which is called point cloud. By the relative positions of the points in space, it is possible to determine what types of objects they belong to and make further decisions based on this information, for example, by constructing a three-dimensional grid of the environment space and filling it with labeled points. The paper considers the problem of segmentation of cloud points - predicting one of the predefined classes to every point. Modern neural network methods allows to solve this problem in a time within 100 ms using GPU and implying the presence of a complete point cloud obtained in one full lidar rotation. At the same time, lidar is able to perform a full turn within the time range of 50 ms to 200 ms depending on the settings. Thereby in some situations the collection of a point cloud can became a bottleneck in the pipeline. In this paper, two approaches are proposed for solving this problem. Both are based on obtaining a 2D representation of the point cloud by projecting collected points onto a sphere, then dividing the projection into disjoint blocks, and further sequentially processing the blocks by neural networks. The first method is based on the modification of existing solutions which use convolutional neural networks. The second approach is based on the convolutional reccurent (ConvLSTM) architecture, which allows to process the point cloud in even smaller parts, considering the data coming from the lidar as a sequence containing spatial information. The paper shows that the first approach could potentially be used to reduce the delay between receiving and using a point cloud. The second approach showed significantly worse quality compared to the modern solutions, however, it allows to solve a simpler binary real-time segmentation problem on CPU, which eliminates the need for the large computing power required for an effective solution to the problem. Theoretically, this allows one to solve the problem on devices with power limitations, for example, small robot couriers or quadrocopters.