UAV's Detection And Classification Of Ground Target

As the interdisciplinary subject of computer vision,artificial intelligence,digital image processing,and human-computer interaction,the object recognition and classification have gained wide attention.With the development of information technology and the increasing popularity of Unmanned Aerial Vehicles(UAV),vision-based drones are playing an cumulatively important role in intelligent video surveillance,3D scene reconstruction,and military applications.This paper focuses on the identification and classification of ground targets by UAV and summarizes the current research status of target detection and image classification at home and abroad,then proposes different methods for diverse tasks.The paper mainly divides into four parts:The first part,there will be the method of object recognition based on multi-feature fusion.This part strats with the introduction of the feature representation methods commonly used in target discernment,in particular,the extraction process of HOG features and LBP feature is explained in detail.Then the PCA algorithm is presented to reduce the dimensionality of the HOG feature,solved the problem that the algorithm is too slow due to the high HOG feature dimension and the HOG-PCA feature is combined with the LBP feature,the SVM classifier is trained using the fused features to achieve the identificcation of the car target in the air-ground scene.The second part,there will be the method of object recognition based on the DPM model.In this section,the calculation process of HOG features in DPM model,the related theoretical knowledge of DPM model and the implementation principle of DPM model for training vehicles using Latent SVM are narrated.In the stage of object recognition,use the trained model to calculate the scores of each position of the scene image,then the NMS removes the duplicated detection frame and the detection frame beyond the image boundary and get a compliant test set.Finally,the location with the largest score in the test set is the position of the target.The third part,there will be the combination of visual saliency and Principal Component Analysis Network(PCANet)for the object recognition.At the beginning of this part,the relevant basics knowledge is expounded,including: K-means algorithm,Mean-Shift algorithm,saliency region detection,convolutional neural network and so on.Among others the the implementation method of training classifier using PCANet is discussed.In the object recognition phase,use the AC saliency detection model to determine the region proposal in the image.Then,the PCANet algorithm is used to obtaine the feature of the target region and loaded into the trained classifier to filter out the non-target areas to achieve detection of tank targets in the scene.The last part,there will be the method of scene image classification and labeling.This part elaborates the image median filtering and scene image segmentation methods.Firstly,Mean-Shift segmentation is performed on the median filtered image,then the contour image is extracted from the segmented image,and the adjacent region of the image is merged by setting the contour threshold to solve the problem of oversegmentation of the Mean-Shift algorithm.Secondly,the color features and texture features of each region are extracted,and the fused features are used as input to the SVM classifier to perform category judgment.Finally,the classification and labeling of scene images are completed.