远距离监视激光雷达动目标快速检测

激光雷达具有全天候工作、探测精度高、有效探测距离远、易获得三维信息等特点，但工作在远距离模式时，目标点云比较稀疏。当前便携条件下，基于深度学习的算法在激光雷达点云数据直接目标识别时，实时性和成功率尚不能达到远程监视实际工程的要求。针对实际工程中利用激光雷达检测运动目标进而实时引导高分辨率相机的需求，采用基于变化的检测方法，对远距离条件下激光雷达的运动目标检测方法进行了研究，利用点云数据的距离信息，给出三维单高斯模型和三维高斯混合模型检测动目标的过程和方法，提出了利用杂波图恒虚警率检测法处理点云数据的方法。实验表明，与二维图像动目标检测方法相比，三维单高斯模型法会很大程度提高检测准确性，降低虚警率，但仍然存在较高虚警率。为适应复杂三维场景，采用基于三维高斯混合模型的方法进一步降低了虚警率，但也降低了检测速度；而杂波图CFAR的方法具有很高的实时性，同时也具有较好的检测性能。

Fast detection of moving targets in long range surveillance LiDAR

  Objective   Lidar is a kind of sensor using laser active imaging, with the advantages of high detection accuracy, all-weather working, easy access to high-precision three-dimensional information, far effective detection range, etc. It has been widely used in recent years, especially in the field of autonomous driving, as a three-dimensional environment perception device in the autonomous driving vehicles. When lidar is applied to perimeter surveillance and working in long-range mode, the target point cloud is relatively sparse which is different from microwave high-resolution imaging radar such as ISAR. The recognition speed of 3D point cloud data with the number of point clouds of 6 000-7 000/frame is lower than 12 frame/s when using training and real-time recognition of cooperative targets by deep learning method, while more missed alarms emerge. The rate of targets recognition needs to be improved. In order to guide the high-resolution infrared camera to carry out high-resolution fine imaging of the detected target before recognition, the method of fast detection of moving targets is investigated. The processing method of complex scenes using 3D Gaussian method and clutter map CFAR to detect moving targets is provided.  Methods   The flow diagram of lidar moving target detection based on 3D point cloud data is given (Fig.2), including 3D point cloud mesh construction, noise filtering by 3D bilateral filtering, target and background segmentation. The principles of 3D single Gaussian method and 3D Gaussian mixture method for segmentation of target/background are given, and the method of using clutter map CFAR detection is proposed (Fig.1). Using 72 frames of data from actual equipment, the result of application of the Faster RCNN Resnet50 FPN deep-learning method, two-dimensional single Gaussian method, three-dimensional single Gaussian method, three-dimensional Gaussian mixture method, and clutter map CFAR method are compared.  Results and Discussions  Comparative experiments show that the average accuracy rate of using the Faster RCNN Resnet50 FPN deep learning model is 0.318 4, the average recall rate is 0.329 4, the processing time of a single frame is 0.5 s, and the point cloud data is 2 s, which means this method is hardware-intensive and difficult to meet the general engineering requirements. In other methods (Tab.2), under the two-dimensional single-Gaussian model, the real-time performance is very high, but there are many false alarms, and almost every frame has false alarms. There are false alarms in some frames of 3D single Gaussian model (Fig.7). By adjusting the parameters of the 3D Gaussian mixture model, the number of false alarms can be reduced to 0 while there are no missed alarms (Fig.8). The false alarm rate will also decrease significantly after using the clutter map CFAR method (Fig.9). At the same time, it can be seen that the processing time of the clutter map CFAR method is basically the same as that of the 3D single Gaussian model method, which is much less than that of the 3D mixed model method, and can meet the actual engineering needs. The 3D Gaussian mixture model needs further optimization or parallel processing to improve real-time performance.  Conclusions   At present, when the deep learning method is directly used to detect and recognize moving targets for the lidar working in the remote monitoring mode, the real-time performance and detection rate can not fully meet the actual engineering requirements. The combination of lidar and high-resolution infrared camera in the project requires lidar to detect moving targets and guide the imaging and recognition of infrared high-resolution camera. Due to the high false alarm rate of two-dimensional single Gaussian method and three-dimensional single Gaussian method, it is difficult to adapt to complex background and cannot meet the requirements. Three-dimensional Gaussian mixture model can adapt to complex background very well, but the real-time performance is reduced because of the increase in the amount of computation caused by the update of background parameters. This means that it can not meet the requirements. In contrast, for the scene with complex background, the method of using clutter map CFAR to detect and process point cloud data can improve the accuracy and the real-time performance of detection, thus meeting the requirements of practical engineering.