Normalized Cross Correlation Research Articles

Endoscopic en-face optical coherence tomography and fluorescence imaging using correlation-based probe tracking.

Forward-viewing endoscopic optical coherence tomography (OCT) provides 3D imaging in vivo, and can be combined with widefield fluorescence imaging by use of a double-clad fiber. However, it is technically challenging to build a high-performance miniaturized 2D scanning system with a large field-of-view. In this paper we demonstrate how a 1D scanning probe, which produces cross-sectional OCT images (B-scans) and 1D fluorescence T-scans, can be transformed into a 2D scanning probe by manual scanning along the second axis. OCT volumes are assembled from the B-scans using speckle decorrelation measurements to estimate the out-of-plane motion along the manual scan direction. Motion within the plane of the B-scans is corrected using image registration by normalized cross correlation. En-face OCT slices and fluorescence images, corrected for probe motion in 3D, can be displayed in real-time during the scan. For a B-scan frame rate of 250 Hz, and an OCT lateral resolution of approximately , the approach can handle out-of-plane motion at speeds of up to 4 mm/s.

Automatic Position Estimation Based on Lidar × Lidar Data for Autonomous Aerial Navigation in the Amazon Forest Region

In this paper we post-process and evaluate the position estimation of pairs of template windows and geo-referenced images generated from LiDAR cloud point data using the Normalized Cross-Correlation (NCC) method. We created intensity, surface and terrain pairs of images for use with template matching, with 5 m pixel spacing, through binning. We evaluated square and circular binning approaches, without filtering the original data. Template matching achieved approximately 7 m root mean square error (RMSE) on intensity and surface templates on the respective geo-referenced images, while on terrain templates it had many mismatches due to insufficient terrain features over the assumed flight transect. Analysis of NCC showed the possibility of rejecting bad matches of intensity and surface templates, but terrain templates required an additional criteria of flatness for rejection. The combined NCC of intensity, surface and terrain proved stable for rejection of bad matches and had the lowest RMSE. Filtering outliers from surface images changed very little the accuracy of the matches, but greatly improved correlation values, indicating that the forest canopy might have the best features for geo-localization with template matching. Position estimation is essential for autonomous navigation of aerial vehicles and the these experiments with LiDAR data show potential for localization over densely forested regions where methods using optical camera data may fail to acquire distinguishable features.

A Fast Template Matching Scheme of Visible and Infrared Image Under Occluded Scenarios

A fast and robust template matching scheme, called Matching by Slice Transform Matrix Mapping (MSTMM), is proposed for the matching difficulty under occluded scenarios caused by nonlinear intensity differences and structure differences between visible and thermal infrared images. The first step in the MSTMM scheme was to extract information about the distribution of pixels with the same gray level through the developed Expanded Slice Transform with Adaptive Gray Level (EST-AGL). After completing the construction of the EST-AGL matrix for all image patches, different EST-AGL matrices were mapped to different integers or floats through the traditional special integer mapping mechanism or the neural network mapping mechanism. Finally, template matching between visible and thermal infrared images was achieved by evaluating the similarity of correlation mapping surface images through the Normalized Cross Correlation (NCC) algorithm. The proposed EST-AGL method can overcome the nonlinear intensity differences between visible and thermal infrared images by extracting the structural features of the image. The mapping mechanism of the MSTMM scheme can reduce the structural differences between the normal template image and the query image under an occluded scenario by increasing the similarity between the normal image patches and the image patches with occlusion. The proper mapping mechanism ensures the high performance of the MSTMM scheme by using only the simple NCC algorithm instead of other time-consuming anti-occlusion dense feature algorithms in the similarity evaluation stage. The three main experimental results of the MSTMM scheme are as follows: (1) the scheme of MSTMM can achieve template matching in only 0.015 seconds when a <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$64\times 64$ </tex-math></inline-formula> template image slides on a <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$256\times 256$ </tex-math></inline-formula> query image on a hardware platform with limited resources; (2) the matching success rate of the MSTMM scheme can reach up to 75% among 2107 experimental samples; and (3) the neural network training in the neural network mapping mechanism only takes at least 104.4 seconds on the CPU.

Object-Based Attention Mechanism for Color Calibration of UAV Remote Sensing Images in Precision Agriculture

Color calibration is a critical step for Unmanned Aerial Vehicles (UAV) remote sensing, especially in precision agriculture, which relies mainly on correlating color changes to specific quality attributes, e.g., plant health, disease, and pest stresses. In UAV remote sensing, the exemplar-based color transfer is popularly used for color calibration, where the automatic search for the semantic correspondences is the key to ensuring the color transfer accuracy. However, the existing attention mechanisms encounter difficulties in building the precise semantic correspondences between the reference image and the target one, in which the normalized cross correlation is often computed for feature reassembling. As a result, the color transfer accuracy is inevitably decreased by the disturbance from the semantically unrelated pixels, leading to semantic mismatch due to the absence of semantic correspondences. In this paper, we proposed an unsupervised object-based attention mechanism (OBAM) to suppress the disturbance of the semantically unrelated pixels, along with a further introduced weight-adjusted AdaIN (WAA) method to tackle the challenges caused by the absence of semantic correspondences. By embedding the proposed modules into a photorealistic style transfer method with progressive stylization, the color transfer accuracy can be improved while better preserving the structural details. We evaluated our approach on the UAV data of different crop types including rice, beans, and cotton. Extensive experiments demonstrate that our proposed method outperforms several state-of-the-art methods. As our approach requires no annotated labels, it can be easily embedded into the off-the-shelf color transfer approaches. Relevant codes and configurations will be available at http://github.com/huanghsheng/object-based-attention-mechanism.

Robust Image-Based Water-Level Estimation Using Single-Camera Monitoring

This study proposes an automatic image water-level measurement system using a single camera. The inverse perspective mapping (IPM) technology is used to rectify an image to improve deviations because the optical axis of the camera and staff gauge are not orthogonal, causing a perspective distortion in the real environment. The initial position of the current watermark is first determined from the staff gauge’s image by applying basic digital image-processing technology. The system uses the image histogram of staff gauge’s region of interest to provide a reference region for determining accurate water levels from the preliminary watermark position. Furthermore, the proposed system can correctly and automatically determine the current water-level elevation by establishing the control points to convert the pixel distance to the real-world distance as well as through the setting of the reference water-level elevation. The images are shifted because of camera vibration, which will also shift the waterline coordinates. To prevent the image shift due to camera vibration from interfering with the accuracy of water-level measurement, we used the normalized cross correlation (NCC) technology with the proposed camera-vibration calibration scheme for removing the impact. The proposed distance calibration (DISCAL) module further optimizes the results of the images rectified using the IPM technology. The suggested automatic water-level measurement system can improve the traditional linear interpolation approach by >90%, and the deviation value of the real water level is < 0.5 cm, which is ideal for water-level measurement in confined spaces according to experimental results.

A rotation invariant template matching algorithm based on Sub-NCC.

This paper proposes an anti-rotation template matching method based on a portion of the whole pixels. To solve the problem that the speed of the original template matching method based on NCC (Normalized cross correlation) is too slow for the rotated image, a template matching method based on Sub-NCC is proposed, which improves the anti-jamming ability of the algorithm. At the same time, in order to improve the matching speed, the rotation invariant edge points are selected from the rotation invariant pixels, and the selected points are used for rough matching to quickly screen out the unmatched areas. The theoretical analysis and experimental results show that the accuracy of this method is more than 95%. For the search map at any angle with the resolution at the level of 300,000 pixel, after selecting the appropriate pyramid series and threshold, the matching time can be controlled to within 0.1 s.

Coregistration of Remote Sensing Image Based on Histogram Kernel Predictability

Registration of remote sensing images has been approached using different strategies; one of the most popular is based on similarity measures. There are different measures of similarity in the literature: Normalized Cross-Correlation (NCC), Mutual Information (MI), etc. Normalized Mutual Information (NMI) has received the most attention in image processing; among the most important limitations are its high computational cost and lack of robustness to strong radiometric changes. For this reason, in this work, we introduce a co-registration approach based on the Histogram Kernel Predictability (HKP). This formulation reduces numerical errors and requires less computing time in comparison to NMI. To the best of our knowledge, this is the first work for registering any remote sensing images by using HKP. Additionally, we propose to use an algorithm based on meta-heuristics called Evolutionary Centers Algorithm (ECA), which allows having fewer iterations to solve the registration problem. In addition, we incorporate a parallelization scheme that permits reducing processing times. The results show that our proposal can solve co-registration problems that the NMI cannot solve while obtaining competitive computational times and registration errors comparable with other existing works in the literature. The HKP approach solves most of all the transformations of a set of simulated registration problems, while the NMI, in some cases, only solves half of the registration problems. Moreover, we compare our approach with feature-based methods in real datasets. This research presents an alternative to remote sensing problems where MI has traditionally been used.

A Hybrid Security Framework for Medical Image Communication

Authentication of the digital image has much attention for the digital revolution. Digital image authentication can be verified with image watermarking and image encryption schemes. These schemes are widely used to protect images against forgery attacks, and they are useful for protecting copyright and rightful ownership. Depending on the desirable applications, several image encryption and watermarking schemes have been proposed to moderate this attention. This framework presents a new scheme that combines a Walsh Hadamard Transform (WHT)-based image watermarking scheme with an image encryption scheme based on Double Random Phase Encoding (DRPE). First, on the sender side, the secret medical image is encrypted using DRPE. Then the encrypted image is watermarking based on WHT. The combination between watermarking and encryption increases the security and robustness of transmitting an image. The performance evaluation of the proposed scheme is obtained by testing Structural Similarity Index (SSIM), Peak Signal-to-Noise Ratio (PSNR), Normalized cross-correlation (NC), and Feature Similarity Index (FSIM).

Toward Computing Cross-Modality Symmetric Non-Rigid Medical Image Registration

This paper describes a new non-rigid approach to register images from same- and cross-imaging modalities such as magnetic resonance imaging, computed tomography, and 3D rotational angiography. The deformation is a key challenge in medical image registration. We have proposed a diffeomorphism-based method to tackle this problem using an optimized framework. A non stationary velocity field is used to minimize the effect of forces that are derived from the image gradients. Furthermore, we propose a similarity energy function, based on the gray scale distribution, to limit the fluctuations while approaching the local minima. The proposed method is evaluated on both private and public datasets; the results show that the values of mean square error (MSE), normalized cross-correlation (NCC), structural similarity (SS), mutual information (MI), feature similarity index (FSIM), and mean absolute error (MAE) are 1.3136, 0.9962, 0.9897, 0.883, 0.9922, and 1.52± 2.09, respectively. Both qualitative and quantitative evaluation show promising registration accuracy reflecting the potential of the proposed method.

Automated Optic Disc Segmentation Using Basis Splines-Based Active Contour

We develop a fully automated method for the segmentation of optic disc in retinal fundus images using basis-spline-based active contour. The segmentation is achieved by performing scaling, translation, and rotation of the active contour, thereby giving rise to five free parameters. The energy of the active contour is defined by the local contrast and is optimized with respect to five free parameters to get the best fit on the optic disc using gradient descent technique and Green&#x2019;s theorem. The use of gradient descent technique and Green&#x2019;s theorem reduces the computational cost and speeds up the segmentation task. The detection of optic disc is achieved using multiresolution-based normalized cross correlation technique. The detection point is used for the initialization of the active contour. Subsequent optimized evolution of the basis-spline-based active contour provides an accurate segmentation of the optic disc. We present validations on the databases such as Drishti-GS, MESSIDOR, RIGA, and a local database containing 101, 1200, 750, and 942 retinal fundus images, respectively, amounting to a total of 2993 fundus images. Their corresponding Dice index scores are 0.9182, 0.8912, 0.9331 and 0.9343. Basic data exploration is done on the results obtained to visualize the trends and distribution of the performance parameters throughout the databases. This also helps us evaluate the algorithm&#x2019;s overall performance more accurately.

Research on the Geometric Precision Correction Method for Three-Axis-Stabilized Geostationary Meteorological Satellite Images Based on Landmarks

To improve the positioning accuracy of the three-axis-stabilized geostationary meteorological satellite (GSMS) images, we developed a geometric precision correction process based on landmarks. For the problem that GSMS images and binary landmark images are difficult to match accurately, we made full use of the local and global spatial information of the image pairs and proposed a dual-threshold matching algorithm based on the misalignment matrix to screen the initial matching results generated by the maximum normalized cross correlation. Under the premise of ensuring matching precision, the number of correctly matched point pairs is effectively increased to provide more sufficient samples for subsequent processing. Finally, we used a shallow neural network to establish the transformation relationship between the GSMS image and the reference image and resampled the GSMS image. We selected observations from five different periods of FY-4A/Advanced Geosynchronous Radiation Imager (AGRI) on-orbit and conducted two verification experiments: 1) in the image matching, compared with several commonly used algorithms, we found that the proposed matching algorithm has obvious advantages in matching precision and relative recall; and 2) in the image correction, the positioning error corrected by our methods is about 1.26 pixels <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$\left ({3\sigma }\right)$ </tex-math></inline-formula> in the central area and about 1.30 pixels <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$\left ({3\sigma }\right)$ </tex-math></inline-formula> in the edge area, which is about 14.9% and 9.1% better than the weighted iterative cubic polynomial fitting algorithm, respectively, with better robustness. In conclusion, the landmark-based geometric precision correction process proposed effectively improves the geometric accuracy of GSMS images, which is an important foundation for subsequent quantitative applications.

Automatic detection and classification of bearded seal vocalizations in the northeastern Chukchi Sea using convolutional neural networks.

Bearded seals vocalizations are often analyzed manually or by using automatic detections that are manually validated. In this work, an automatic detection and classification system (DCS) based on convolutional neural networks (CNNs) is proposed. Bearded seal sounds were year-round recorded by four spatially separated receivers on the Chukchi Continental Slope in Alaska in 2016-2017. The DCS is divided in two sections. First, regions of interest (ROI) containing possible bearded seal vocalizations are found by using the two-dimensional normalized cross correlation of the measured spectrogram and a representative template of two main calls of interest. Second, CNNs are used to validate and classify the ROIs among several possible classes. The CNNs are trained on 80% of the ROIs manually labeled from one of the four spatially separated recorders. When validating on the remaining 20%, the CNNs show an accuracy above 95.5%. To assess the generalization performance of the networks, the CNNs are tested on the remaining recorders, located at different positions, with a precision above 89.2% for the main class of the two types of calls. The proposed technique reduces the laborious task of manual inspection prone to inconstant bias and possible errors in detections.

Wavelet thresholding techniques implementation in retinal images for suppressing noises

In medical image processing, noise plays a major role of reducing the details of the images and blurs the features which are important for the diagnosis of the disease. Retinal images are high resolution images. Gaussian noise, Salt & Pepper noise affects certain important features of them which are needed for proper diagnosis of the retinal diseases. So removing noises in these images is a kind of difficulty that all researcher experiences. The main objective is to reduce the effect of Gaussian noise and Salt & Pepper noise in retinal images. Recently, wavelet transform is being used in all areas of image processing and especially in medical image denoising, it is highly used because it has multi resolution property and sparsity. So here, wavelet thresholding techniques such as Visu shrink, SURE Shrink and Bayes shrink are implemented to remove Gaussian noise and Salt & Pepper noise. The performance of all the three methods are evaluated using PSNR (Peak Signal to Noise Ratio), MSE (Mean Square Error), NK (Normalized cross correlation) and Time elapsed. Also, the performances of all the thresholding techniques were compared using above mentioned metrics.

Towards efficient network compression via Few-Shot Slimming

While previous network compression methods achieve great success, most of them rely on the abundant training data which is, unfortunately, often unavailable in practice due to some reasons, e.g., privacy issues, storage constraints, and transmission limitations. A promising way to solve this problem is to perform compression with a few unlabeled data. Proceeding along this way, we propose a novel few-shot network compression framework named Few-Shot Slimming (FSS). FSS follows the student/teacher paradigm, and contains two steps: (1) construct the student by inheriting principal feature maps from the teacher; (2) refine the student feature representation by knowledge distillation with an enhanced mixing data augmentation method called GridMix. Specifically, in the first step, we employ normalized cross correlation to perform the principal feature analysis, and then theoretically construct a new indicator to select the most informative feature maps from the teacher for the student. The indicator is based on the variances of feature maps which can efficiently quantitate the information richness of the input feature maps in a feature-agnostic manner. In the second step, we perform the knowledge distillation for the initialized student in first step with a novel grid-based mixing data augmentation technique which greatly extends the limited sample dataset. In this way, the student is able to refine its feature representation and achieves a better result. Extensive experiments on multiple benchmarks demonstrate the state-of-the-art performance of FSS. For example, by using 0.2% label-free data of full training set, FSS yields a 60% FLOPs reduction for DenseNet-40 on CIFAR-10 with only a loss of 0.8% in top-1 accuracy, achieving a result on par with that obtained by the conventional full-data methods.

Detrended multiple moving average cross-correlation analysis and its application in the correlation measurement of stock market in Shanghai, Shenzhen, and Hong Kong

Based on the traditional detrended moving average cross-correlation analysis (DMCA) method and the detrended multiple cross-correlation analysis (DMC) method, we propose the detrended multiple moving average cross-correlation analysis method (DMMC) to quantify the correlation between non-stationary signals on different time scales. It will be the generalization of the correlation between pairwise time series and can be used to measure the multivariate correlation between the dependent variable y with the others i independent variable x. To validate the DMMC method, we determined the DMMC coefficient into numerical simulations on the standard Gaussian distribution, Lognormal distribution t distribution and ARFIMA long memory random sequences. Then we apply the proposed method to explore the multiple correlations between China’s Shanghai, Shenzhen, and Hong Kong stock markets and analyze the dynamic time-varying characteristics of the correlation with the sliding window approach. The results show that DMMC is suitable to measure multiple cross-correlations. This new method can be further applied to multiple cross-correlation in different areas of knowledge, such as energy, meteorology, medicine, transportation, etc.

Cryptocurrency Market Consolidation in 2020-2021.

Time series of price returns for 80 of the most liquid cryptocurrencies listed on Binance are investigated for the presence of detrended cross-correlations. A spectral analysis of the detrended correlation matrix and a topological analysis of the minimal spanning trees calculated based on this matrix are applied for different positions of a moving window. The cryptocurrencies become more strongly cross-correlated among themselves than they used to be before. The average cross-correlations increase with time on a specific time scale in a way that resembles the Epps effect amplification when going from past to present. The minimal spanning trees also change their topology and, for the short time scales, they become more centralized with increasing maximum node degrees, while for the long time scales they become more distributed, but also more correlated at the same time. Apart from the inter-market dependencies, the detrended cross-correlations between the cryptocurrency market and some traditional markets, like the stock markets, commodity markets, and Forex, are also analyzed. The cryptocurrency market shows higher levels of cross-correlations with the other markets during the same turbulent periods, in which it is strongly cross-correlated itself.

Development and evaluation of an unbiased, high‐resolution and detail‐preserving structural T1‐weighted template for use in studies on older adults

AbstractBackgroundConstruction of an optimal T1‐weighted template for older adults often requires selection of a target anatomy for nonlinear‐registrations, however a poorly chosen template may introduce bias due to large structural discrepancies between template and subjects used for template construction resulting in misregistration and thereby loss of important anatomical details. Moreover, large structural brain differences across older adults may introduce registration errors. Here, a minimum‐deformation T1‐weighted template specifically for studies on older‐adults was developed using a sparse, patch‐based, detail‐preserving template construction method that minimizes biases.MethodThis work used T1‐weighted brain MRI data obtained from 400 non‐demented older adults (65‐95 years of age, male:female=1:1) participating in longitudinal cohort studies of aging. The template construction scheme involved a reference‐free groupwise normalization utilizing affine transformations followed by ANTs SyGN groupwise normalization utilizing a weighted average scheme in every iteration. Finally a patch‐based tissue‐guided sparse‐representation approach was employed to generate the new template (Fig. 1). The new template was compared to the weighted average template and other 0.5mm templates through visual inspection, in terms of image sharpness by means of the normalized power spectral density and also inter‐subject spatial normalization accuracy achieved when used as references for normalization of T1‐weighted data from 202 ADNI3 participants. Normalization accuracy was assessed for each template by means of the average pairwise normalized cross correlation, average pairwise Jaccard index and average log‐Jacobian determinant of ADNI3 participants.ResultThe new template maintained anatomical consistency and exhibited higher image sharpness as demonstrated by larger high spatial frequency content in the normalized power spectra compared to other templates (Fig. 2 &amp; Fig. 3A,C). Detailed cortical features were better resolved with the new template compared to MCALT_0.5mm (Fig. 3A,B). Visual inspection revealed atypical anatomical structures in the cortex of MCALT_0.5mm (red circles) (Fig. 3B), absent in other templates. Inter‐subject spatial normalization accuracy was higher when using the new template compared to the other templates (Fig. 4A,B,C). The new template required lower deformation for spatially normalizing ADNI3 dataset (Fig. 5).ConclusionThe new template is an unbiased high‐resolution detail‐preserving structural template of the older adult brain that requires lower deformation and provides higher spatial normalization accuracy for older adult data than other templates.

Image Representation Method Based on Complex Wavelet Transform and Phase Congruency, with Automatic Threshold Selection

Image representation is an active area of research with increasing applications in military and defense. Image representation aims at representing an image with lesser number of coefficients than the actual image, without affecting the image quality. It is the first step in image compression. Once the image is represented by using some set of coefficients, it is further encoded using various compression algorithms. This paper proposes an adaptive method for image representation, which uses Complex Wavelet transform and the concept of phase congruency, where the number of coefficients used for image representation depends on the information content in the input image. The efficiency of the proposed method has been assessed by comparing the number of coefficients used to represent the image using the proposed method with that used when Complex Wavelet transform is used for image representation. The resultant image quality is determined by computing the PSNR values and Normalized Cross Correlation. Experiments carried out show highly promising results, in terms of the reduction in the number of coefficients used for image representation and the quality of the resultant image.

A Physics-Informed Neural Network Approach for Nearfield Acoustic Holography.

In this manuscript, we describe a novel methodology for nearfield acoustic holography (NAH). The proposed technique is based on convolutional neural networks, with autoencoder architecture, to reconstruct the pressure and velocity fields on the surface of the vibrating structure using the sampled pressure soundfield on the holographic plane as input. The loss function used for training the network is based on a combination of two components. The first component is the error in the reconstructed velocity. The second component is the error between the sound pressure on the holographic plane and its estimate obtained from forward propagating the pressure and velocity fields on the structure through the Kirchhoff–Helmholtz integral; thus, bringing some knowledge about the physics of the process under study into the estimation algorithm. Due to the explicit presence of the Kirchhoff–Helmholtz integral in the loss function, we name the proposed technique the Kirchhoff–Helmholtz-based convolutional neural network, KHCNN. KHCNN has been tested on two large datasets of rectangular plates and violin shells. Results show that it attains very good accuracy, with a gain in the NMSE of the estimated velocity field that can top dB, with respect to state-of-the-art techniques. The same trend is observed if the normalized cross correlation is used as a metric.

Long-Term Target Tracking of UAVs Based on Kernelized Correlation Filter

During the target tracking process of unmanned aerial vehicles (UAVs), the target may disappear from view or be fully occluded by other objects, resulting in tracking failure. Therefore, determining how to identify tracking failure and re-detect the target is the key to the long-term target tracking of UAVs. Kernelized correlation filter (KCF) has been very popular for its satisfactory speed and accuracy since it was proposed. It is very suitable for UAV target tracking systems with high real-time requirements. However, it cannot detect tracking failure, so it is not suitable for long-term target tracking. Based on the above research, we propose an improved KCF to match long-term target tracking requirements. Firstly, we introduce a confidence mechanism to evaluate the target tracking results to determine the status of target tracking. Secondly, the tracking model update strategy is designed to make the model suffer from less background information interference, thereby improving the robustness of the algorithm. Finally, the Normalized Cross Correlation (NCC) template matching is used to make a regional proposal first, and then the tracking model is used for target re-detection. Then, we successfully apply the algorithm to the UAV system. The system uses binocular cameras to estimate the target position accurately, and we design a control method to keep the target in the UAV’s field of view. Our algorithm has achieved the best results in both short-term and long-term evaluations of experiments on tracking benchmarks, which proves that the algorithm is superior to the baseline algorithm and has quite good performance. Outdoor experiments show that the developed UAV system can achieve long-term, autonomous target tracking.