Research On Data Processing And Classification Of Harvesting Targets In Plantation Forest Environment

Environmental information measuring system was built by laser scanner, visual equipment and vehicle-mounted holder and the effective multi-source information of several sensors was collected in order to solve the problems of operation interruption, low efficiency and the risk of driver caused by the complex environment of forestry scene and uncertain classification of targets during the driving and harvesting process of special forestry equipment in plantation forest environment. Then super-voxel process and step-by-step cascading segmentation were provided to handle fusion point cloud. Finally, a new pattern recognition algorithm was proposed to classify harvesting targets and barriers in plantation forest environment, which improved the automatic and intelligent level of special forestry equipment effectively. The main research work are listed as the following:1. According to the environmental characteristics of special forestry equipment’s operation, the vehicle-mounted holder was designed to carry two dimensionality laser range finder, inertial measurement unit and camera with high depth of field to build the real-time acquisition system. The visual and three dimensional space information of target object in the scene was real-time collected and storage through the host control software, this software also was used to set up and monitor the working state of vehicle-mounted holder and the measuring units.2. Filtering process was made to remove the random and systems noise of 3D laser point cloud. The internal parameters of two sensors were calibrated by analyzing the working principle of laser scanner and the camera. Then a multi-segment external reference parameter calibration model was constructed by designing an octagon calibration board, whose edges have corresponding relationships between 3-D multiple point cloud coordinate system and pixel coordinate system. After compensating jitter error and step angle error of the laser point cloud, the external calibration parameters optimized by nonlinear method were acquired, and the effective integration of 3D laser and visual data was finished.3. For the beam enlarging and energy loss problem during the process of propagation and return, the spot size and energy distribution characteristics were applied on the process of voxelization. And the super-voxel’s central point was confirmed on the basis of the gradient of voxel size. Then determine the optimum central point and divide the field by iterative optimization clustering method. With the feature extracted, the fusion point clouds were turned into super-voxel.4. In the super-voxel feature vector space, step-by-step cascade segmentation combined by ground integrated feature distinct segmentation, optimization Gaussian peak clustering segmentation and CRFs segmentation was applied to segment independent objects such as harvesting targets and barriers self-adaptively in acquired plantation forestry environment. This method made up for "uncertain area" segmentation issues.5. The features including color, shape, reflected energy and spatial distribution were extracted for the segmented harvesting targets and barriers in plantation forestry environment on the basis of image and 3d laser point clouds data. Then the multiclass gaussian granular kernel fuzzy support vector machine model was proposed to classify those targets successfully. By comparison with other models, the classification model in this article is more effective for harvesting target identification with high performance, the correct recognition rate could reach 98.2%.