Research On Harvesting Targets Identification In Forest Based On 2D Laser And Images

In order to solve the problems of low efficiency and high danger for forest harvester operation caused by the forest environment, terrain, and presence of obstacles, the information collection system for forest environment was built based on 2D laser scanner and infrared thermal imager. Then harvesting targets could be detected and identified by using information fusion technology and be shown in picture, which provided supplementary information to the operator to compensate for the lack of the human eyes. This method reduced the risks in operation and laid the theory and technology foundation for the realization of the intelligent automation forestry equipment.The main research work and conclusions are listed as the following:1. In accordance with the characteristics of the forest environment and the actual test conditions, 2D laser rangefinder and infrared cameras were selected to collect forest environment information, where the PC was applied to control the information collection. Also the collected signal pretreatment was conducted in this part.2. According to the working principle of the two sensors, internal and external joint calibration was conducted between laser points and image to match them together, thereby obtaining the target area image we need at the same time to obtain the coordinate position of the target by the laser information. Further optimization for calibration parameters was done, and then the more precise target area in images could be acquired to provide the basis for subsequent target recognition.3. Infrared image were fused with visible image fusion, which was achieved by PCA and PCNN respectively. Then features selection and extraction were conducted based on images, which offered daba basis for subsequent research.4. Based on the 150 collected trees, pedestrians and rock sample data set in the earlier stage, popular machine learning algorithms, including AdaBoost, K neighbors, artificial neural networks and support vector machine, were used to carry out samples classification. For neural network algorithm, four models with the best classification performance were established by a combination of different training functions.5. SVM was selected for further study. Based on 500 samples, model optimization was carried out in two ways. On the one hand, SVM internal parameters were optimized by different optimization algorithms; on the other hand the fuzzy SVM model was constructed and a new fuzzy membership calculation method was proposed. Based on different databases, the algorithm was verified and the results showed that the proposed new algorithm could improve SVM performance with a high recognition rate at 96%.