All-day Real-time Monitoring Of Slope Rockfall Based On Semantic Segmentation And LiDAR Laser Point Cloud

Slope failure is a type of sudden and highly destructive natural disaster,with wideranging hazards and difficulty in prevention and forecast.Disasters result not only in casualties but also in damage to surrounding buildings and economic losses.In some cases,they can also pose a threat to the ecological environment.Therefore,real-time monitoring of slope failure is crucial.Remote sensing and ground monitoring are currently the two main methods for monitoring slope failure,but these methods suffer from problems such as poor real-time performance,strong dependence on light sources,and difficulty in installation.Prior to a large-scale outbreak of slope failure,there are often precursors such as falling rocks and soil loosening in small areas.However,these methods are limited by the resolution of remote sensing images and sensors,and cannot effectively identify the early signs of slope failure.Aiming at the task of real-time monitoring of slope failure without light source,thesis proposes a real-time monitoring method for slope failure using Li DAR with semantic segmentation.This research is outlined as follows:(1)Design and construct a simulation experimental platform for slope failure.Use the MANTIS VISION F6 Smart scanner to capture 3D point cloud data of slope and slope disasters in environment with and without light source,and add indexing to each point in the point cloud to create a dataset of slope failure point clouds.(2)A slope area division method is proposed.Point cloud data collected under lighting conditions are subjected to orthogonal projection transformation to generate a twodimensional image of the slope.The Fast-SCNN fast semantic segmentation network is employed to partition the slope area and identify the boundary that separates the slope area from the human activity area.The slope failure monitoring method proposed in this paper only requires a light source for the experiments in this chapter.Subsequent work can be completed without a light source.(3)A non-light source,real-time slope rockfall failure detection method is proposed.Point-Net network is used to segment the rockfall from the point cloud of the slope background,achieving real-time detection of slope disasters.(4)Trajectory prediction and warning of rockfall.Firstly,according to the segmentation results in(3),the coordinates of the center point of the rockfall are obtained,and a motion model of the rockfall is established using the Kalman filter to achieve trajectory prediction of the rockfall.Secondly,establish a mapping relationship between the three-dimensional point cloud and the two-dimensional image.Project the slope boundary information obtained from(2)onto the three-dimensional point cloud and compare it with the disaster trajectory to determine whether the disaster spreading area has reached the warning threshold.According to the experimental results,in the task of slope area division,the segmentation accuracy exceeds 97%;in the task of real-time detection of slope rockfall failure,the segmentation accuracy exceeds 98%,and an average of 28 sets of point cloud data could be processed per second.In the task of rockfall trajectory prediction,the average error was only 9.67 mm.In summary,the proposed method in this paper can achieve real-time monitoring of slope failure under non-light source conditions.