LiDAR Point Cloud Data Augmentation Under the Influence of Large-Scale Smoke Diffusion in Autonomous Driving

With the rapid development of smart cities and Internet of Things (IoT) technologies, the application of IoT sensing in smart firefighting and disaster response has gradually become a research and practical focus. As the primary sensing method for IoT sensors, LiDAR's performance under the influence of smoke has become particularly important. So far, no dataset with smoke characteristics is available. The actual LiDAR dataset for recording large-scale fires is difficult to operate, dangerous, and costly. In this regard, this article uses a smoke physics diffusion model to simulate the smoke environment at the fire scene on existing datasets. The resulting enhanced dataset is used to train the network to reuse these datasets, aiming to achieve a more robust fire smoke neural network. The smoke physics diffusion model is parameterized, and wind direction and wind speed are added. Finally, a smoke diffusion model under the intensity and time of the diffusion source is obtained. The Hausdorff distance, Chamfer distance (CD), and Earth mover's distance (EMD) of the smoke point cloud are used as evaluation criteria, and the actual smoke data and simulated smoke data are compared to verify the effectiveness and accuracy of the above model. Combined with the LiDAR sensor model, the mechanism of smoke affecting LiDAR sensor imaging is studied. Based on this mechanism, an LiDAR smoke impact model with the Rayleigh scattering model as the core is established, and the smoke diffusion model is combined with LiDAR data. This model modifies the smoke-free dataset to obtain a smoke dataset enhanced by the smoke model that conforms to the sensor characteristics and actual smoke distribution. In order to verify the effectiveness and universality of the enhanced dataset, data training is performed on point cloud deep learning networks such as point-based RandLA-Net, grid-based Point-Voxel-KD, and projection-based Range-Net++ to verify the effectiveness of the above work. Finally, the training results show that the data enhanced by smoke conform to the general law of the influence of smoke on object detection in the actual objective world. The presence of smoke will reduce the accuracy of network recognition, which has a greater impact on the projection-based point cloud deep learning method.