Pedestrian Detection and Tracking in Sparse Mls Point Clouds Using A Neural Network and Voting-Based Approach

This paper presents and extends an approach for the detection of pedestrians in unstructured point clouds resulting from single MLS (mobile laser scanning) scans. The approach is based on a neural network and a subsequent voting process. The neural network processes point clouds subdivided into local point neighborhoods. The member points of these neighborhoods are directly processed by the network, hence a conversion in a structured representation of the data is not needed. The network also uses meta information of the neighborhoods themselves to improve the results, like their distance to the ground plane. It decides if the neighborhood is part of an object of interest and estimates the center of said object. This information is then used in a voting process. By searching for maxima in the voting space, the discrimination between an actual object and incorrectly classified neighborhoods is made. Since a single labeled object can be subdivided into multiple local neighborhoods, we are able to train the neural network with comparatively low amounts of labeled data. Considerations are made to deal with the varying and sparse point density that is typical for single MLS scans. We supplement the detection with a 3D tracking which, although straightforward, allows us to deal with objects which are occluded for short periods of time to improve the quality of the results. Overall, our approach performs reasonably well for the detection and tracking of pedestrians in single MLS scans as long as the local point density is not too low. Given the LiDAR sensor we used, this is the case up to distances of 22 m.