Research On Classification Technology Of Streak Tube Lidar Echo Data Based On Random Forest

As one of the Li DAR with broad research prospects,the Streak Tube Imaging Li DAR(STIL)can obtain high-precision three-dimensional information of multiple detection targets as the same time on the basis of full-waveform sampling,it has broad development prospects in airborne lidar mapping.In the real-time lidar mapping process,the echo data obtained by the fringe tube lidar system per second can reach several hundred megabytes.The analysis and processing of high-speed massive echo data have become the current key research direction.The existing classification and recognition method are to classify and identify the point cloud data reversed by the fringe echo data.However,the existing classification and recognition method are to classify and identify the point cloud data that are reversely derived from fringe echo data.This secondary data processing method reduces the accuracy and timeliness of radar data and has a certain impact on the application and promotion of radar in the later stage.Therefore,it is necessary to study the direct processing of the fringe tube lidar echo data.With the development of machine learning and deep learning,more and more machine learning methods are introduced into the field of image processing to solve tasks such as image recognition and image classification.Among them,the random forest algorithm is a strong classifier based on the decision tree algorithm and integrated learning method.It performs well in both classification tasks and regression tasks.Based on the random forest algorithm,this paper proposes to directly classify the original fringe echo signals and combines the image recognition technology to propose a method for individually extracting and classifying the target area in the fringe image.This method combines the rapidity and anti-jamming ability of random forest algorithm,solves the problem of poor timeliness and accuracy of existing fringe echo data secondary classification method,and has good adaptability in the process of different terrain classification.In this experiment,according to the theory of morphological characteristics and local features,combined with the feature extraction process,we have established the stripe image feature extraction algorithm,the characteristics of the original fringe echo signal are analyzed and extracted.A random forest classification model is established by combining the feature set and the training process of the model,the classification accuracy of the initial model is 86.6%.After the completion of the model,use the grid-search method to adjust the number of base decision trees and the maximum depth of the base decision tree of the random forest model.The number of base decision trees of the optimal random forest classification model is 130 and the maximum base decision tree depth Is 8,the classification accuracy rate is 89.3%,and the sample feature importance is calculated by the OOB score of the optimal model,thereby verified the feature distribution in the feature database.By comparing the accuracy and recall rate of the random forest algorithm and other classification algorithms when performing the two and three classification tasks of streak echo signals,it can be seen that the random forest algorithm has a significant advantage in processing streak echo data.Finally,a classification model is used to classify and mark the echo signal data.The final classification results are converted into point cloud and displayed in point cloud,and typical targets are extracted and converted into point cloud.