Semantic Segmentation and Construction of a DataSet From a 3D Point Cloud Obtained by LiDAR Sensor

The use of LiDAR technologies offer an effective alternative for enhancing monitoring and surveillance tasks in secured areas, or hard-to-access areas. Conventional solutions allow detection and identification capabilities, which are valuable for object and pattern recognition in environments with special conditions. This paper proposes a method for structure a dataset by collecting data from LiDAR sensors and reconstructing 3D point cloud scenes into images. Furthermore, training with these data using semantic segmentation algorithms allows for the detecting of people in monitoring and surveillance contexts. The process involves data acquisition techniques, preprocessing, and the application of tools and technologies for semantic segmentation in the post-processing stage. Data is acquired from RPLidar sensor with a spatial resolution of less than 0.4 degrees per step, ensuring greater accuracy in detecting small objects. Scanning distances between 0.15 and 12 meters allow for the capture of 941 points per rotation angle. The dataset is complemented with labels and images containing binary masks of regions of interest for people detection using semantic segmentation. The main contribution of this paper is the methodological description of constructing a dataset comprising 1,254 images in various scenarios, including indoors and outdoors, with different levels of occlusion and illumination. This dataset facilitates the development of machine learning algorithms and deep learning algorithms for pattern recognition, such as Mask R-CNN (89.9% mean class accuracy) and U-NET (90.5% mean class accuracy). The results demonstrated an approximate 2% improvement in mean class accuracy compared to other state-of-the-art algorithms in semantic segmentation for point clouds.