Auto-Learning Correlation-Filter-Based Target State Estimation for Real-Time UAV Tracking

Abstract

Most existing tracking methods based on discriminative correlation filters (DCFs) update the tracker every frame with a fixed learning rate. However, constantly adjusting the tracker can hardly handle the fickle target appearance in UAV tracking (e.g., undergoing partial occlusion, illumination variation, or deformation). To mitigate this, we propose a novel auto-learning correlation filter for UAV tracking, which fully exploits valuable information behind response maps for adaptive feedback updating. Concretely, we first introduce a principled target state estimation (TSE) criterion to reveal the confidence level of the tracking results. We suggest an auto-learning strategy with the TSE metric to update the tracker with adaptive learning rates. Based on the target state estimation, we further developed an innovative lost-and-found strategy to recognize and handle temporal target missing. Finally, we incorporated the TSE regularization term into the DCF objective function, which by alternating optimization iterations can efficiently solve without much computational cost. Extensive experiments on four widely-used UAV benchmarks have demonstrated the superiority of the proposed method compared to both DCF and deep-based trackers. Notably, ALCF achieved state-of-the-art performance on several benchmarks while running over 50 FPS on a single CPU. Code will be released soon.

1. Introduction

Given a target object specified by a bounding box in the first frame, visual object tracking aims to determine the target’s exact location sequentially in a video, which serves as a fundamental task in the computer vision community. Recently, increasing attention has been paid to unmanned aerial vehicle (UAV) tracking [1,2,3,4,5] due to its promising development in the field of geomatics [6], agroforestry [7], transportation [8] and mid-air aircraft tracking [9].

Existing object tracking methods can mainly be divided into two types: discriminative correlation filter (DCF)-based [10,11,12,13,14] and deep-based [15,16,17,18]. Since UAVs only provide limited power capacity and computational resources, deep-based trackers usually suffer low inference speeds and high computational overheads. In contrast, DCF-based trackers achieve an ideal trade-off between accuracy and speed by efficiently computing the Fourier frequency domain, providing good solutions to real-time UAV tracking.

DCF-based trackers [3,11,12,13] detect the target by generating a response map and assigning the peak position as the target position. Then, the model will be updated with a fixed learning rate after every frame to adapt to the target’s appearance change. However, due to the common challenges of fast motion and out-of-view in UAV tracking, target appearance could change drastically, rending the ineffectiveness of a stable update strategy. An appropriate update strategy is a goal to pursue. Recently, some trackers [2,19,20,21] have tried to mine information from response maps, such as fluctuation degree and peak height, as updated guidance. Intuitively, the peak height can provide a rough indication of the similarity between the target and the tracker, and the fluctuation degree indicates background clutter to some extent. LMCF [19] proposed an criterion called average peak-to-correlation energy (APCE), beyond which a model update is required. We regard it as a good measure of the fluctuation degree, but this single criterion seems not robust enough to cope with the appearance change, as shown in Figure 1. It is not sensitive to slight target state changes (e.g., in frames 1300, 1400), and there is an unnecessarily sharp drop when the model state is credible (in frame 1268). In addition, although LMCF stops the model from updating in some frames, which can prevent model degradation to a certain extent, it still relies on the pre-set single learning rate to cope with the model updating.

1.1. Related Works

1.1.1. Discriminative Correlation Filter

Discriminative correlation filter (DCF)-based methods have been widely applied to visual tracking due to their high computational efficiency. Bolme et al. [22] proposed a minimum output sum of squared error (MOSSE) filter and were the first to apply a correlation filter to object tracking. Henriques et al. [23] introduced a circulant matrix to produce sufficient samples for training and detection while maintaining a fast tracking speed. However, the resulting periodic repetitions at boundary positions limit the discrimination of the tracker. To mitigate this issue, various regularization terms have been introduced [2,12,24]. Danelljan et al. [12] adopted a spatial regularization term that penalizes the filter coefficients for background regions to force the tracker to focus more on the target center. Huang et al. [3] used response maps as regularization terms for model training and repressed the aberrance of response maps to improve tracking performance. Chen et al. [4] designed a Gaussian-like response map as expected output to enhance the representation learning ability of the correlation filter. Xu et al. [25] proposed to gauge the relevance of multi-channel features for the purpose of channel selection to improve the performance of DCF in accuracy and provide a parsimonious model from the attribute perspective. Wang et al. [26] proposed a context and saliency-aware CF to strengthen the ability to extract targets of interest from complex background. Fu et al. [27] introduced an object-saliency-aware, dual-regularized correlation filter to suppress the boundary effect and consequently enhance the performance of the tracker. Li et al. [28] introduced a temporal regularization term, which uses the model trained in the previous frame as the reference model to constrain the model training in the current frame, thereby preventing model degradation. However, it uses a fixed constant as the regularization coefficient and adopts the model in the previous frame as the reference model regardless of whether the target state in the previous frame is credible.

1.1.2. Prior Knowledge to Model Update

Recently, increasing attention has been paid to the model updating. Wang et al. [19] introduced a criterion called APCE to measure response map quality and guide model updates. Li et al. [2] used the variations of response maps between consecutive frames to estimate the credibility of the target state and restricted model update when needed. Huang et al. [29] updated the Siamese tracker by transferring the previous target knowledge with attention. Dai et al. [30] used an offline-trained meta-updater to determine whether or not to update the tracker with two fixed learning rates (0 or 1), which fails to adapt to fickle target appearances varying at different rates in UAV tracking scenarios. Though adopting adaptive learning rates, Yang et al. [31] highly relies on LSTM [32], causing a high computational cost. Chen et al. [33] constructed a Gaussian-like (GL) function label for correlation filter training and proposed an accurate incremental update to mitigate model degradation by combining target samples with adaptive aspect ratios. In this paper, we fully exploit the information of the response map and directly obtain the learning rate from it, which is computationally efficient and suitable for real-time tracking. To the best of our knowledge, no previous works have exploited the quality of response maps to evaluate the target state and adjust the model’s learning rate adaptively.

1.1.3. Redetect the Lost Target

Due to the complex situations during tracking, targets may be missed in some frames. To address this problem, previous works [34,35,36] often designed an additional detector to search for the target on the entire screen after target loss. Some trackers [37,38] even introduce global detection into tracking and locate the target through spatio-temporal constraints. However, these methods are challenging to run in real-time on GPU, which limits their applications in UAV tracking. Huang et al. [9] exploited motion features of a moving target to redetect the target from clutter background after loss, but this rough approach cannot handle static targets or visible videos. Zhu et al. [21] expanded the search area to increase the possibility of recapturing the target. Recently, Alan Lukezic et al. [36] verified the possibility of correlation filters as a detector using multiple correlation filters updated at different time scales. However, the large number of filters used in [36] slows the tracking speed, which limits its application in the field of UAV. Therefore, exploring an efficient re-detection method for UAV tracking is very meaningful. Unlike previous works, we introduce an iterative search method with only one correlation filter, which tracks the target and searches for the target after it is lost, achieving tracking performance improvement at the expense of only a slight decrease in tracking speed.

1.2. Contributions

We present a novel auto-learning correlation filter (short for ALCF) tracking algorithm for efficient UAV tracking. We first introduce an innovative criterion, target state estimation (TSE), to comprehensively get aware of the target state. Considering the peak height and fluctuation degree, both informative features of response maps, simultaneously, the resulting TSE value can represent the uncertainty in the target state of each frame reliably, thereby having a keener sense of slight target state changes and being more resistant to noise. As shown in Figure 1, TSE remains steady when the target moves smoothly and fluctuates significantly when occlusion or target lost occurs (as denoted by the circles). Our TSE successfully indicates whether the target is currently in the presence of occlusions (green circle), partially lost (yellow circle), or out-of-view (blue circle). It is noted that the comparison of APCE and TSE in Figure 1 was performed on the same tracker (ALCF).

Based on the TSE, we propose a novel auto-learning strategy that automatically derives a rapid and optimal learning rate from TSE adaptively depending on the estimated target states. Specifically, when the TSE remains large, indicating a transparent and credible target state, no update for the tracker is required, since it nicely fits the target. Moreover, when the TSE decreases slightly, indicating the target undergoes deformation, a significant learning rate is produced to adapt to rapid appearance changes. Furthermore, when the TSE drops sharply and the target state is uncertain, the target may suffer severe occlusion or be lost from view such that a small learning rate is generated to retain prior knowledge of the target and avoid model degradation. This way, ALCF can adjust its tracker properly under complex and changeable situations.

Moreover, the target in UAV tracking usually suffers large positional displacement between successive frames or even temporal disappearance from view, which quickly results in tracking failures. The challenges lie in recapturing the target when it appears in subsequent frames, as large positional displacement may exist. Thanks to the TSE criterion, which would decrease sharply when the target is lost (as exampled by the blue circle in Figure 1), we present a new lost-and-found strategy that perceives such a TSE decrease and enlarging search areas centered at the location, where the target last appears immediately. ALCF can quickly recapture the lost target with a new alternate search strategy with negligible extra computational cost.

The proposed TSE can also contribute to model optimization. We introduce a novel regularization term by using the trusted model trained with previous frames as a reference model to prevent model degradation. Instead of manually setting the coefficient for the regularization term, ALCF can automatically adjust the coefficient according to TSE. When the target state is credible, the coefficients are small to relax the model variation, and vice versa—the coefficients are significant in constraining the variation of the model in the current frame. We also incorporate two additional regularization terms to help suppress the aberration of computed responses for consecutive frames [3] and focus more on the central portion of targets [12].

We extensively evaluated our ALCF on several commonly used benchmark datasets, including UAVDT [39], UAV123@10fps [40], DTB70 [41], and VisDrone2018-test-dev [42], which contain large-scale, challenging aerial video sequences captured by UAVs. We provide two version trackers of ALCF, one based on a simple scale regression strategy (proposed by DSST [43]), which far exceeds real-time on a single CPU and performs favorably against other state-of-the-art trackers with hand-drafted features. Another version of ALCF exploits deep features for scale regression (proposed by Alpha-Refine [44]), which shows competitive performance against recent deep trackers.

To sum up, this work makes the following contributions:

• We present a novel TSE metric to the community which allows accurate estimation of the target state.
• We introduce a new ALCF tracker upon TSE, which consists of a novel auto-learning strategy, a fast lost-and-found strategy, and an effective regularization term for efficient and robust UAV tracking.
• We propose a deep version model which outperforms numerous recently popular deep framework-based trackers and established the new state-of-the-art on several UAV datasets.
• We provide the community with an optimal hand-drafted feature-correlation filter at over 50 FPS on a single CPU.

2. Method

In this section, we first describe the tracking mechanism of the correlation filter and the overall framework of the ALCF tracker, followed by a detailed description of the construction process of target state estimation (TSE), and finally describe the three strategies based on TSE.

2.1. Overview

The overall framework of our method is shown in Figure 2. The DCF-based tracker consists of three stages: the detection stage for target localization, the online model update stage for learning target appearance information, and the training stage for computing correlation filters. Based on the TSE value computed from the response map, we propose a lost-and-found strategy in the detection stage to cope with target loss. Then, an auto-learning strategy to learn new information from the target sample is presented, which learns adaptively according to the target state. Finally, a model that constrains regularization is introduced into model training to obtain a robust correlation filter model for tracking.

In the detection stage, the correlation filter $f_{t-1}$ from the previous frame is used to search for the target in the current frame. The correlation response map is used to locate the target position, and the peak position is considered the target position of the current frame. Besides the peak position of the response map, we also exploit other information (i.e., peak value, fluctuation degree) to calculate TSE to evaluate the target state. When there is a sudden drop in the TSE value, we judge that the target has been lost and start using the lost-and-found strategy to re-detect the target, skipping the online update stage and the training stage if the target is lost.

For the online update stage, we propose an auto-learning strategy to learn new information from the current image. Unlike traditional DCF-based trackers that use a pre-set constant as the learning rate, ALCF uses TSE to compute an adaptively changing learning rate $\eta$, which allows the appearance model to learn more from the current image when the target state is credible and less when the target state is in doubt, thus maintaining robust tracking.

For the training phase, we present the model constrain regularization strategy to train the model. Specifically, we use the reference model $f_r$ preserved from historical frames to prevent model degradation and use TSE to guide the update of the reference model and its weight coefficient $\gamma$ during model training. When the target state is credible, we reduce the weight coefficient of the reference model and relax the restriction on model learning from the current frame, in the meanwhile updating the reference model to use the newly computed correlation filter $f_t$ as the reference model for model training in subsequent frames. Conversely, we increase the weight coefficient of the reference model and restrict the model changes to prevent model degradation. Then, the model can be solved efficiently via the alternating direction method of multipliers (ADMM) algorithm.

2.2. Target State Estimation (TSE)

We developed TSE by comprehensively evaluating the peak value $F_{max}$ and fluctuation degree (short for $FD$) of response maps, which can indicate the uncertainty in the target state and thus reflect the target state.

Denote F as the response map of dimensions $W\times H$ (width × height in pixels). We first compute mean-square error (MSE) to measure the difference between F and the ideal Gaussian response:

$$MSE=\frac{1}{WH}·\sum _{i=1}^W\sum _{j=1}^H{(\mathit{F}(i,j)-\mathit{y}(i,j))}^2,$$

(1)

where $F(i,j)$ represents the response value on pixel $(i,j)$, y is the ideal response map (Gaussian distribution), and F is expected to be similar to y when the target state is credible, leading to a low MSE value.

Given MSE, the TSE is computed as:

$$TSE={[1+e^{-(T-\mu )}]}^{-1},$$

(2)

$$T={\alpha }_1{F}_{max}+{\alpha }_{2}FD,$$

(3)

$$FD=\frac{{F_{max}}^2}{MSE},$$

(4)

where $F_{max}$ is the peak value of F, which measures the similarity score for the tracker and the target; $FD$ reflects the fluctuation degree of F and the degree of background disturbance. When there are no distractors (i.e., frames 7, 21, and 307 in Figure 3), the value of $FD$ is about 20 times that of $F_{max}$. Thus, we use two constant coefficients, ${\alpha }_1$ and ${\alpha }_2$, to ensure $F_{max}$ and $FD$ are of the same magnitude. $\mu$ is an offset constant. When the target state is credible, $F_{max}$ is large while $MSE$ is small, resulting in a considerable TSE value (close to 1). In contrast, TSE would drop to 0 when the target is lost. When the target suffers deformation (in frame 21, the target suffers out-of-plane rotation), the peak value drops slightly. When there exist distractors (in frame 228), sub-peaks lead to a significant drop to $FD$, but $F_{max}$ is still high, keeping the TSE value enormous. Thus, with the combined effect of peak and fluctuation degree, our TSE is more robust and better guides the target state changes. More details about the construction of TSE can be found in Section 3.3.5.

2.3. Auto-Learning Correlation Filter

Based on TSE, we developed a new ALCF tracker comprised of an auto-learning strategy, a lost-and-found strategy, and an effective regularization term.

2.3.1. Auto-Learning Strategy

Existing methods update the tracking model with a single learning rate as follows:

$${\widehat{\mathit{x}}}_k^{model}=(1-\eta ){\widehat{\mathit{x}}}_{k-1}^{model}+\eta {\widehat{\mathit{x}}}_k,$$

(5)

where ^ denotes the discrete fourier transform (DFT) of a signal, ${\widehat{x}}_k$ is the vectorized image of frame k, and ${\widehat{x}}_{k-1}^{model}$ is the tracking model learned from frame $k-1$. $\eta$ is a fixed learning rate that is hard to set manually for UAV tracking, as constant and uncertain appearance changes exist. For example, when the correlation filter model can describe the target well, the model does not need to learn new information from the target sample image. When the target suffers severe occlusion or is lost from the screen, the model should stop updating to prevent contamination by untrustworthy samples. On the other hand, when there exists a difference between the model and the current target appearance, the model should accelerate learning new information from the target appearance. When there exists a huge difference between the target appearance and the model, we argue that maybe the target is lost, and the learning rate should reduce to keep the prior knowledge of the target.

Given this, we developed an auto-learning strategy that updates the tracking model with varying learning rates depending on the target state indicated by TSE, as shown in Figure 3. We automatically adjust the learning rate depending on different target states revealed by TSE computed from the distribution of response maps. When the target moves smoothly (frame 12), TSE is close to one, and the learning rate drops to zero, since updating the tracker is not necessary. However, when the target suffers occlusion or deformation (frames 32, 48, and 156), TSE decreases while the learning rate increases to adapt to rapid changes in target appearance. Moreover, when the target is lost from view (frame 20), TSE drops sharply, and the learning rate becomes zero to stop the model updating and retain the prior knowledge of the target. Concretely, we use TSE to compute the learning rate of frame k from a Gaussian distribution:

$${\eta }_k=\frac{1}{\sqrt{2\pi }\sigma }{e}^{-\frac{{(TSE-\beta )}^2}{2{\sigma }^2}}-\phi ,$$

(6)

where $\beta$ is set as $0.5$, since TSE is bounded between 0 and 1, the standard deviations $\sigma$ is set as 1, and $\phi$ is a constant to make ${\eta }_k$ fall in a reasonable value range. When the value of TSE is close to 1, the target state is trusted, and the learning rate is low; and as the value of TSE drops, the learning rate takes the lead and then decreases. When TSE decreases to 0, we think the target may have been lost and stopped updating.

2.3.2. Lost-and-Found Strategy

We designed a novel lost-and-found strategy based on TSE. When the target is lost from view, TSE will drop suddenly. For example, it decreases by over seventy percent within ten frames, as illustrated in Figure 4. ALCF keeps a close eye on the trend of TSE and initiates the search for the lost target around the location where it last appeared (green rectangle) upon a sharp drop in the TSE. TSE drops rapidly when the target suffers complete occlusion, as shown by the TSE curve. ALCF initiates four additional search windows (yellow dashed box) and iterates their space intervals over three increased distances frame by frame to enlarge search areas. Until any one of the search windows regains a large TSE (more extensive than the sum value of the rest windows, as shown in frame 49), the tracker successfully recaptures the lost target and terminates the re-found process.

Since large positional displacement may exist when the lost target reappears in subsequent frames, previous works have introduced additional detectors to search for the target in the entire frame, leading to heavy memory and computation cost. Instead, we increase the number of search windows to cover a large search area and thus avoid using a detector. However, simply increasing search windows, though increasing the chances of recapturing the target, would hurt tracking speed at the same time. Instead of augmenting massive search windows in every single frame, we explored an efficient alternating search scheme that gradually enlarges the distance of four additional search windows in consecutive frames to cover large search areas sequentially, as shown by the yellow dashed rectangle in Figure 4. We set three distance intervals based on the width w and height h of the target and iterate over them every three consecutive frames until the target is re-found.

Our lost-and-found strategy exploits the correlation filter as the detector, which has two advantages: (i) The re-detection process is efficiently computed in the Fourier domain. The alternate search mechanism does not introduce a significant additional computational burden, which is friendly to the limited computational resources of UAVs. (ii) The correlation filter can be used not only as a tracking model but also as a detector when the target is lost, which brings no additional memory pressure and has low memory requirements for the UAVs.

2.3.3. Model Constrain Regularization

We further introduced a new regularization term by employing the prior knowledge learned from previous frames to constrain the filter learning of the current frame:

$$R_1=\frac{{\gamma }_k}{2}{∥{\mathit{f}}_k-{\mathit{f}}_r∥}_2^2,$$

(7)

where $f_k\in {\mathbb{R}}^T$ denotes the filter of frame k, and $f_r\in {\mathbb{R}}^T$ represents the reference filter learned from previous frames, which encodes informative prior knowledge about the target. Instead of updating $f_r$ after each frame [28], we only use the filter learned from the frame with favorable target states (e.g., no severe occlusion and deformation) as a reliable reference filter. To this end, we use TSE to estimate the target state of every frame and only use the filter from the frame whose computed TSE is larger than a pre-set threshold $\tau$ to refresh $f_r$.

In addition, we use an adaptive coefficient ${\gamma }_k$ to adjust the weight of the regularization term depending on the current target state:

$${\gamma }_k=\frac{1}{2}(1-TS{E}^2),$$

(8)

when the current target state is trustworthy, i.e., TSE is large, ${\gamma }_k$ is small such that $f_k$ learns less from $f_r$ to adapt to the latest appearance change of the target. Otherwise, $f_k$ borrows more knowledge from $f_r$ to prevent model degradation.

2.4. Model Optimization

The objective function can be formulated as:

$$\begin{array}{c}\hfill E\left(f_k\right)=\frac{1}{2}{∥\sum _{d=1}^D({\mathit{x}}_k^d\ast {\mathit{f}}_k^d)-\mathit{y}∥}_2^2+\sum _{i=[1,2,3]}{R}_i,\end{array}$$

(9)

where $x_k^d\in {\mathbb{R}}^T$ denotes the d-$th$ channel of the vectorized image of frame k, and D is the total channel number. $y\in R^T$ is the expected response (Gaussian distribution). ∗ indicates the convolution operator. $R_{i\in \{1,2,3\}}$ are three additional regularization terms to improve the discrimination of the model, which are detailed below.

$R_1$ is the regularization term defined in Equation (7). We further employ the spatial regularization term [12] to eliminate boundary effects caused by circulant shifted samples:

$$R_2=\frac{1}{2}\sum _{d=1}^D{∥\mathit{w}\odot {\mathit{f}}_k^d∥}_2^2,$$

(10)

where w is a negative Gaussian-shaped spatial weight vector to make the tracking model emphasize more on the target center. ⊙ indicates the Hadamard product operation.

In addition, a regularization term for response maps [3] is also used to exploit informative features of response distribution:

$$R_3=\frac{\theta }{2}{∥\sum _{d=1}^D({\mathit{x}}_k^d\ast {\mathit{f}}_k^d)-\sum _{d=1}^D({\mathit{x}}_{k-1}^d\ast {\mathit{f}}_{k-1}^d)\left[{\Psi }_{p,q}\right]∥}_2^2,$$

(11)

where $\theta$ is a regularization coefficient; $x_{k-1},{\mathit{f}}_{k-1}\in {\mathbb{R}}^T$ denote the vectorized image and the filter of frame $k-1$, respectively; ${\sum }_{d=1}^D({\mathit{x}}_{k-1}^d\ast {\mathit{f}}_{k-1}^d)$ is the computed response map. p and q denote the location differences of two peaks in both response maps in two-dimensional space, and $\left[{\Psi }_{p,q}\right]$ is used to align the peak position of two consecutive response maps.

2.5. Optimization Solutions

For the ease of optimization, we introduce an auxiliary variable ${\widehat{g}}_k=\sqrt{N}\mathit{F}{\mathit{f}}_k$, where F is the orthonormal $N\times N$ matrix of complex basis vectors for mapping any N dimensional vectorized signal to Fourier frequency domain. Thus, we transformed the objective function into the frequency domain:

$$\begin{array}{cc}\hfill & E({\mathit{f}}_k,{\widehat{\mathit{g}}}_k^d)=\frac{1}{2}{∥\sum _{d=1}^D({\widehat{\mathit{x}}}_k^d\ast {\widehat{\mathit{g}}}_k^d)-\widehat{\mathit{y}}∥}_2^2+\frac{1}{2}\sum _{d=1}^D{∥\mathit{w}\odot {\mathit{f}}_k^d∥}_2^2\hfill \\ \hfill & +\frac{\theta }{2}{∥\sum _{d=1}^D({\widehat{\mathit{x}}}_k^d\ast {\widehat{\mathit{g}}}_k^d)-{\mathit{R}}_{pre}∥}_2^2+\frac{{\gamma }_k}{2}\sum _{d=1}^D{∥{\widehat{\mathit{g}}}_k^d-{\widehat{\mathit{g}}}_r^d∥}_2^2,\hfill \end{array}$$

(12)

where $R_{pre}$ represents the discrete Fourier transformation of the previous frame’s response map, which is a constant for the current frame. We used the alternating direction method of multipliers (ADMM) algorithm to minimize Equation (12) to achieve a locally optimal solution. The augmented Lagrangian form of the equation can be formulated as:

$$\begin{array}{cc}\hfill L_k({\mathit{f}}_k,{\widehat{\mathit{g}}}_k^d,{\mathit{\xi }}^T)& =E({\mathit{f}}_k,{\widehat{\mathit{g}}}_k^d)+{\widehat{\mathit{\xi }}}^T({\widehat{\mathit{g}}}_k-\sqrt{N}\mathit{F}{\mathit{f}}_k)\hfill \\ \hfill & +\frac{\rho }{2}{∥{\widehat{\mathit{g}}}_k-\sqrt{N}\mathit{F}{\mathit{f}}_k∥}_2^2,\hfill \end{array}$$

(13)

where ${\widehat{\xi }}^T=[{\widehat{\mathit{\xi }}}_1^T,{\widehat{\mathit{\xi }}}_2^T,\dots ,{\widehat{\mathit{\xi }}}_D^T]$ is the $1\times DN$ Lagrangian vector in the Fourier domain, $\rho$ is the penalty factor, and superscript T indicates the conjugate transpose operation. Then, ADMM is applied by alternately solving the following two sub-problems:

2.5.1. The Solution to Sub-Problem ${\widehat{g}}_{k+1}$

The sub-problem ${\widehat{g}}_{k+1}$ can be formulated as:

$$\begin{array}{cc}\hfill {\widehat{\mathit{g}}}_{k+1}& =argmin\frac{1}{2}{∥\sum _{d=1}^D({{\widehat{\mathit{x}}}_k}^d\odot {{\widehat{\mathit{g}}}_k}^d)-\widehat{\mathit{y}}∥}_2^2\hfill \\ \hfill & +\frac{{\gamma }_k}{2}\sum _{d=1}^D{∥{{\widehat{\mathit{g}}}_k}^d-{{\widehat{\mathit{g}}}_r}^d∥}_2^2+\frac{\theta }{2}{∥\sum _{d=1}^D({{\widehat{\mathit{x}}}_k}^d\odot {{\widehat{\mathit{g}}}_k}^d)-{\mathit{R}}_{pre}∥}_2^2\hfill \\ & +{\widehat{\mathit{\xi }}}^T({\widehat{\mathit{g}}}_k-\sqrt{N}\mathit{F}{\mathit{f}}_k)+\frac{\rho }{2}{∥{\widehat{\mathit{g}}}_k-\sqrt{N}\mathit{F}{\mathit{f}}_k∥}_2^2,\hfill \end{array}$$

(14)

which is too complex to solve directly. Given the fact that $x_k$ is sparse-banded, and thus each element of $\widehat{y}(\widehat{\mathit{y}}\left(t\right),t=1,2,\dots ,N)$ is dependent only on each ${\mathit{x}}_k\left(t\right)={[x_k^1\left(t\right),x_k^2\left(t\right),\dots ,x_k^D\left(t\right)]}^T$ and ${\widehat{g}}_k\left(t\right)={[conj({\widehat{g}}_k^1(t)),conj({\widehat{g}}_k^2(t)),\dots ,conj({\widehat{g}}_k^D)(t)]}^T$ [45], $conj(.)$ denotes the complex conjugate operation. Thus, we divided the vectorized image ${\mathit{x}}_k$ into N elements and solved the problem by N smaller and independent problems, over $t=[1,2,\dots ,N]$:

$$\begin{array}{cc}\hfill & {\widehat{\mathit{g}}}_{k+1}\left(t\right)=argmin\frac{1}{2}{∥{{\widehat{\mathit{x}}}_k}^T\left(t\right)\odot {\widehat{\mathit{g}}}_k\left(t\right)-\widehat{\mathit{y}}\left(t\right)∥}_2^2\hfill \\ \hfill & +\frac{{\gamma }_k}{2}{∥{\widehat{\mathit{g}}}_k\left(t\right)-{\widehat{\mathit{g}}}_r∥}_2^2+\frac{\theta }{2}{∥({{\widehat{\mathit{x}}}_k}^T\left(t\right)\odot {\widehat{\mathit{g}}}_k\left(t\right))-{\mathit{R}}_{pre}∥}_2^2\hfill \\ \hfill & +{\widehat{\mathit{\xi }}}^T({\widehat{\mathit{g}}}_k\left(t\right)-\sqrt{N}\mathit{F}{\mathit{f}}_k\left(t\right))+\frac{\rho }{2}{∥{\widehat{\mathit{g}}}_k\left(t\right)-\sqrt{N}\mathit{F}{\mathit{f}}_k\left(t\right)∥}_2^2,\hfill \end{array}$$

(15)

where $f_k\left(t\right)=[f_k^1\left(t\right),f_k^2\left(t\right),\dots f_k^D\left(t\right)]$, ${\widehat{x}}_k^T\left(t\right)$ is the discrete Fourier transformation of $x_k^T\left(t\right)$. Each smaller problem can be calculated efficiently to achieve the solution of Equation (15):

$${\widehat{\mathit{g}}}_{k+1}\left(t\right)={((1+N\theta ){\widehat{\mathit{x}}}_k\left(t\right){\widehat{\mathit{x}}}_k^T\left(t\right)+(\rho +{\gamma }_k)N)}^{-1}\mathit{H},$$

(16)

$$\mathit{H}={\widehat{\mathit{x}}}_k^T\left(t\right)\widehat{\mathit{y}}\left(t\right)+N\theta {\mathit{R}}_{pre}{\widehat{\mathit{x}}}_k^T\left(t\right)+N{\widehat{\mathit{g}}}_r-N{\mathit{\xi }}^T+\rho N{\widehat{\mathit{f}}}_k.$$

(17)

We use Sherman–Morrison formula [46] to accelerate the inverse operation:

$${(M+u{v}^T)}^{-1}=M^{-1}-M^{-1}u{v}^T{M}^{-1}(1+v^T{M}^{-1}u).$$

(18)

Hence, Equation (16) can be reformulated as:

$${\widehat{\mathit{g}}}_{k+1}\left(t\right)=\frac{1}{(\rho +{\gamma }_k)N}(I-\frac{{\widehat{\mathit{x}}}_k\left(t\right){\widehat{\mathit{x}}}_k^T\left(t\right)}{\frac{(\rho +{\gamma }_k)N}{1+\theta N}+{\widehat{\mathit{x}}}_k\left(t\right){\widehat{\mathit{x}}}_k^T\left(t\right)})\mathit{H}.$$

(19)

2.5.2. The Solution to Sub-Problem $f_k$

The subproblem $f_k$ can be written as:

$$\begin{array}{cc}\hfill {\mathit{f}}_{k+1}& =argmin\frac{1}{2}\sum _{d=1}^D{∥\mathit{w}\odot {\mathit{f}}_k^d∥}_2^2\hfill \\ \hfill & +{\widehat{\mathit{\xi }}}^T({\widehat{\mathit{g}}}_k-\sqrt{N}\mathit{F}{\mathit{f}}_k)+\frac{\rho }{2}{∥{\widehat{\mathit{g}}}_k-\sqrt{N}\mathit{F}{\mathit{f}}_k∥}_2^2,\hfill \end{array}$$

(20)

which can be easily solved as follows:

$${\mathit{f}}_{k+1}=\frac{{\mathit{\xi }}^{T}N+\rho N{\mathit{g}}_{k+1}}{\mathit{w}\odot \mathit{w}+\rho N}.$$

(21)

2.5.3. Update of Lagrangian Parameter

We update Lagrangian multipliers as:

$${\mathit{\xi }}^{i+1}={\mathit{\xi }}^i+\rho ({\mathit{g}}_{k+1}^{i+1}-{\mathit{f}}_{k+1}^{i+1}),$$

(22)

where i denotes the iteration number. $g_{k+1}^{i+1}$ and $f_{k+1}^{i+1}$ indicate the solution of subproblem $g_{k+1}$ and $f_{k+1}$, respectively. The coefficient $\rho$ (initially equals 1) is computed as:

$${\rho }^{i+1}=min({\rho }_{max},\beta {\rho }^i),$$

(23)

where $\beta$ and ${\rho }_{max}$ are set as 10 and 10,000, respectively.

3. Results and Discussion

In this section, we first compare our ALCF with state-of-the-art trackers with four challenging UAV datasets, then provide experimental analysis to validate our design choices. Finally, a deep version of ALCF is proposed to compare with recent deep trackers to validate the superior performance of ALCF.

3.1. Implementation Details

All experiments were performed using MATLAB R2019a on a computer with Intel i7-9700 [email protected] GHz, 16 GB RAM, and NVIDIA GTX 1080. We set $\theta =2.2, \tau =0.9,$$\phi =0.35$, $\mu =6$. Considering the numerical magnitude difference between $F_{max}$ and $FD$, we set ${\alpha }_2=20\times {\alpha }_1=500$. The iteration number of ADMM was 2. We extracted histogram of oriented gradients (HOG) [47] and color name (CN) [48] features for input images. A separate filter in [43] was used in our ALCF for scale estimation, and the scale refine module proposed in [44] was used in DeepALCF for scale estimation. Note that the parameters of ALCF remained fixed on all image sequences on all benchmarks. ALCF and DeepALCF had the same parameters except for the size estimation method.

Two widely-used metrics, the precision plot and success plot, describe the performance of all trackers through one-pass evaluation (OPE). Concretely, a precision plot computed the percentages of frames in which the estimated target location is within a given distance threshold from the ground truth. We used the score at the threshold of 20 pixels for ranking the trackers. The success plot measures the fractions of successful frames in which the intersection over union (IoU) between the predicted bounding box and the ground truth is more significant than a certain threshold varied from 0 to 1. We use the area under curve (AUC) of the success plots to rank all trackers.

3.2. Comparison with Hand-Crafted Trackers

We compared our ALCF with thirteen state-of-the-art hand-crafted trackers, including AutoTrack [2], STRCF [28], ECO [49], CSR-DCF [10], MCCT-H [50], ARCF [3], BACF [13], DSST [43], fDSST [11], SRDCF [12], Staple [51], SAMF [52], and KCF [53]. The detailed comparisons on different benchmarks are listed below. Note that the publicly available codes and default parameters offered by the authors were utilized for impartial comparison. In Figure 5, Figure 6, Figure 7 and Figure 8, the numbers in the legend indicate the average precision scores for the precision plot and average AUC scores for the success plot.

3.2.1. Results on UAV123@10fps

UAV123@10fps [40] consists of 123 challenging sequences with 12 different attributes. All sequences were temporally down-sampled to 10 FPS to simulate challenging large target displacements between consecutive frames. Figure 5 provides comparisons with other preeminent trackers. ALCF achieved the best performance in both precision and success. Notably, ALCF outperformed the previous state-of-the-art tracker, AutoTrack, by a large margin, i.e., 1.8% and 1.5%, in precision and success, respectively. This evidences that ALCF can cope well with sudden changes in target appearance and large position shifts between adjacent frames, which commonly appear in UAV tracking.

3.2.2. Results on VisDrone2018-Test-Dev

VisDrone2018-test-dev [42] contains 35 sequences captured by various UAV platforms in over 14 different cites in China, featuring diverse real-world scenarios. Figure 6 shows that our ALCF obtained the best success score of 0.585 and a competitive precision score of 0.798, which is on par with the best-performing tracker ECO, mainly due to the auto-learning and the lost-and-found strategy in ALCF. The auto-learning strategy can adjust the learning rate according to the target state to prevent model degradation. The lost-and-found strategy can help the tracker recapture target after the target is lost, enhancing the performance in coping with long-term tracking (the average length of VisDrone2018-test-dev is around 940 frames; the most extended sequence can reach a length of 2569 frames).

3.2.3. Results on DTB70

DTB70 [41] comprises 70 sequences covering different types of UAV movements, including rapid translation and rotation. Targets include humans, animals, and rigid objects. In Figure 7, the numbers in the legend indicate the average precision scores for the precision plot and average AUC scores for the success plot. Figure 7 depicts that our ALCF performed the best in terms of success rate and second-best in precision. We attribute this to the auto-learning update strategy in ALCF, since there are many viewpoint changes in DTB70, where auto-learning plays an important role.

3.2.4. Results on UAVDT

UAVDT [39] was selected from over 10 h of videos taken by a UAVs at different locations in urban areas. The targets are mainly vehicles with additional attributions, i.e., weather conditions, flying altitude, and camera view. As we can see in Figure 8, our tracker is the best in terms of success. It outperformed the previous best tracker, ECO [49], and second-best tracker, AutoTrack [2], by 0.5% and 0.6%, respectively.

3.2.5. Average Performance Results

Figure 9 reports the average precision and success rate of the 13 trackers and ALCF on the four benchmarks. ALCF yielded the best average success rate score and precision score. It is noted that ALCF was the only tracker with a success rate score (0.503) above 0.5, surpassing the second-best tracker AutoTracker (0.494) by 0.9%. As far as we know, ALCF is the best tracker with hand-drafted features on the four benchmarks.

3.2.6. Per-Attribute Evaluation

Figure 10 provides the precision comparison of twelve attributes defined on UAV123@ 10fps. The number in parentheses is the precision of ALCF. Detailed definitions of these attributions can be found on UAV123@10fps. Our method performed the best against state-of-the-art trackers (i.e., AutoTrack [2], ARCF [3], BACF [13], ECO [49], and STRCF [28]) in most attributes. It is worth noting that our ALCF performed extremely well not only in coping with rapid changes (e.g., viewpoint or illumination variation, and camera motion), but also in preventing model degradation (e.g., out-of-view, partial and full occlusion). These challenges abound in real-world UAV object tracking, which is ample evidence that our strategies are quite effective for UAV tracking.

3.2.7. Visualization

Figure 11 illustrates some qualitative comparisons of ALCF with other top-performing trackers on some challenging sequences (from top down they are ChasingDrones (DTB70), StreetBasketball3 (DTB70), bike2 (UAV123@10fps), person10 (UAV123@10fps),and S0602 (UAVDT)). Our ALCF showed superiority over other trackers in handling challenging scenarios, including fast motion (the first sequence), heavy occlusion (the second sequence), low resolution (the third sequence), loss from view (the fourth sequence), and viewpoint change (the last sequence). Of course, our algorithm still has limitations. In the case of long-term occlusion and superimposition of target deformation, tracking loss will occur.

3.2.8. Speed

Table 1. Average speed (FPS) of top five trackers on [39,40,41,42], ranked by average success (top 3 in bold).

	ALCF	AutoTrack	ECO-HC	STRCF-HC	ARCF
FPS	51.3	58.3	69.4	23.2	23.3
Success	0.503	0.494	0.487	0.469	0.469
Precision	0.724	0.723	0.696	0.672	0.697

compares the tracking speeds of ALCF and other DCF-based trackers (top-5 trackers with hand-drafted features in terms of success over four benchmarks). All the trackers can run on a single CPU. ALCF reached a real-time speed of 51.3 FPS, comparable to that of ECO and AutoTrack, with better predictive performance.

3.3. Ablation Studies

We investigated the effectiveness of different components of ALCF on UAV123@10fps. As presented in

Table 2. Ablation analysis. R represents the response map regularization term. AUC represents the success rate (top 1 in bold).

	AL	MC	LF	R	Precision	AUC	FPS
Baseline					0.600	0.432	36.94
	✓				0.637	0.460	45.64
		✓			0.649	0.463	34.03
			✓		0.648	0.465	32.38
		✓	✓		0.654	0.468	32.36
	✓	✓	✓		0.667	0.473	47.89
	✓			✓	0.654	0.469	42.44
		✓		✓	0.650	0.464	33.70
			✓	✓	0.652	0.474	36.52
		✓	✓	✓	0.662	0.475	29.91
	✓	✓	✓	✓	0.689	0.492	41.31
Baseline+STRCF					0.638	0.457	31.87
Baseline+MC+Fix					0.639	0.459	33.40

, we used the model variant, which employs the same objective function and fixes the learning rate of 0.025 as in SRDCF [12] as a baseline. For a fair comparison, the baseline model adopted the same features extraction and scale estimation scheme as ALCF (original SRDCF utilizes only HOG features, so the performance in Figure 5 is not as good as that in Table 2). AL, MC, and LF denote the auto-learning strategy, model constraint regularization term, and lost-and-found strategy, respectively. Different components are incrementally added to the baseline model individually or in combination to verify their effectiveness.

3.3.1. Effect of Auto-Learning Strategy

As shown in Table 2, Baseline+AL outperformed the baseline in both precision and success rate by 3.7% and 2.8%, respectively. In addition, incorporating an auto-learning strategy improves tracking speed by 23.6%, as it helps reduce meaningless or detrimental training of individual samples. The tracking speed in Table 2 is lower than the average speed in Table 1 because the targets in UAV123@10fps vary drastically between consecutive frames compared to those in the other three datasets, resulting in the frequent model updating in most frames. Nevertheless, our auto-learning strategy still improved the tracking speed significantly.

3.3.2. Effect of Regularization Term

Baseline+MC denotes the model variant that incorporated the proposed regularization term into the baseline model, Baseline+MC+Fix replaced the adaptive regularization coefficient $\gamma$ calculated from Equation (8) with a fixed constant, and Baseline+STRCF updated the reference model in Equation (7) after each frame without considering target states [28]. As is shown in Table 2, using a credible model as a reference and an adaptive regularization coefficient contributes more to the performance boost.

3.3.3. Effect of the Lost-and-Found Strategy

We present a model variant (Baseline+LF) with a lost-and-found strategy added to the baseline model. Notably, employing the lost-and-found strategy boosted the performance significantly by 4.8% and 3.3% in precision and success rate, respectively, along with a negligible increase in tracking speed.

3.3.4. Effect of the Objective Function

We also found that all three components can be further improved to different degrees with the help of the additional response map regularization term used in [3], mainly attributed to its aberration suppression of the response maps between two adjacent frames, which helps the response maps to be more stable. The TSE proved to be more accurate (the performance of Baseline+AL+MC+LF+R (precision: 0.689, AUC: 0.492, FPS: 41.31) is higher than that of Baseline+AL+MC+LF (precision: 0.667, AUC: 0.473, FPS: 47.89) with a slight sacrifice in speed). We also verified the generality of our three innovations, which can improve the tracking performance for trackers with different objective functions.

3.3.5. Construction of Target State Estimation

In our experiment, we found that T (calculated from Equation (3)) has an extensive variation range, which exceeded 10 when the target state was trustworthy and dropped below 2 when the target was lost. We should focus on the variation when T falls between 2 to 10 to perceive the transition of target states. Hence, we introduce the sigmoid activation function (with $\mu$ as 6) upon T to magnify its variation in the interval [2,10] while suppressing noisy fluctuations when the target state is entirely credible ($T>10$) or completely untrustworthy ($T<2$).

We also experimented on UAV123@10fps to validate our parameter settings. Figure 12 shows the performance peaks at $\mu =6$. When $\mu$ becomes small, the value of TSE becomes large (close to 1), which leads to a gradual reduction of the learning rate to 0. When $\mu =0$, it is equivalent to not updating the model, so the model’s performance was the worst. Moreover, when $\mu$ becomes large, the value of TSE will become smaller, which caused the model to use a large learning rate to update when it did not need to be updated and zero when it needed to be updated. However, the performance was slightly better compared to the case of not updating the model ($\mu =0$), which verified the importance of the model update.

3.3.6. Sensitivity Analysis of Model Constraint Regularization Term

Based on STRCF [28], we further designed an adaptive model constraint regularization term for UAV tracking, which can update the reference model adaptively and adjust the effect of the reference model on the correlation filter learning. Moreover, we validated our parameter settings on UAV123@10fps. We set a threshold $\tau$ to refresh $f_r$. When the value of TSE is larger than $\tau$, the reference model $f_r$ will be updated. To find the optimal value of $\tau$, we refined the search within the range of values of $\tau$, taking values at 0.1 intervals. The result is shown in Figure 13. We can see that when $\tau =0$, the reference filter $f_r$ is updated every frame with an adaptive parameter $\gamma$. When $\tau =1$, the reference filter $f_r$ is never updated during tracking. By comprehensively considering the precision and success rate, we set $\tau =0.9$.

3.4. Extension to Deep Trackers

We further evaluated the tracking performance of ALCF in comparison with other state-of-the-art deep-based trackers on the four benchmarks. Inspired by [44], we proposed a deep version of ALCF called DeepALCF, which exploits Alpha-Refine as our scale estimation. Note that DeepALCF still exploits HOG and CN as feature presentations for object location, but it exploits ResNet-34 to extract deep features for scale estimation, which is more accurate. More detail about the scale estimation can be found in Alpha-Refine [44].

3.4.1. Results on VisDrone2018-Test-Dev

We selected 10 new state-of-the-art deep trackers for comparison. The results are shown in

Table 3. Performance comparison of ALCF with other ten state-of-the-art deep trackers (top 3 in bold).

	MDNet [16]	UDT+ [15]	PrDiMP [57]	ATOM [20]	DiMP [55]	SiamRPN [17]	CFNet [58]	Super_DiMP [56]	KYS [54]	HiFT [59]	ALCF Ours	DeepALCF Ours
Precision	0.790	0.800	0.797	0.751	0.806	0.781	0.778	0.800	0.810	0.721	0.798	0.816
Success	0.579	0.584	0.600	0.563	0.605	0.573	0.568	0.610	0.600	0.527	0.585	0.619

. DeepALCF yielded the best precision score (0.816), surpassing the second-best tracker KYS [54] and third-best tracker DiMP [55] by 0.6% and 1%. DeepALCF also obtained the best success score (0.619), outperforming the second-best tracker, Super_DiMP [56] (0.610), and the third-best tracker, DiMP [55] (0.605) by 0.9% and 1.4%. This proves that the auto-learning strategy can cope well with object tracking in natural scenes.

3.4.2. Results on UAVDT

Table 4. Overall evaluation of ALCF, DeepALCF, and some other state-of-the-art trackers on UAVDT [39] (top 3 in bold).

Trackers	Venue	Precision	Success
MDNet [16]	CVPR2016	0.725	0.464
MCCT [50]	CVPR2018	0.691	0.448
ASRCF [24]	CVPR2019	0.720	0.449
TADT [60]	CVPR2019	0.700	0.441
UDT [15]	CVPR2019	0.674	0.442
UDT+ [15]	CVPR2020	0.696	0.415
PrDiMP [57]	CVPR2020	0.757	0.559
fECO [61]	TIP2020	0.699	0.415
fDSTRCF [61]	TIP2020	0.677	0.454
HiFT [59]	ICCV2021	0.652	0.474
LUDT [62]	IJCV2021	0.631	0.418
LUDT+ [62]	IJCV2021	0.701	0.406
ALCF	Ours	0.725	0.456
DeepALCF	Ours	0.748	0.534

shows the precision and success scores of ALCF and other 12 deep trackers on UAVDT. ALCF was very competitive in comparison with PrDiMP [57] in both precision and success and outperformed third-best tracker MDNet [16] by 2.3% and 7% in terms of precision and success, respectively.

3.4.3. Results on DTB70

We compared DeepALCF with 13 popular deep trackers.

Table 5. Overall evaluation of ALCF and some other state-of-the-art trackers on DTB70 [41] (top 3 in bold).

Trackers	Venue	Precision	Success
HCF [63]	ICCV2015	0.616	0.415
MDNet [16]	CVPR2016	0.703	0.466
CFNet [58]	CVPR2017	0.587	0.398
IBCCF [64]	ICCVW2017	0.669	0.460
CREST [65]	ICCV2017	0.650	0.452
ADNet [66]	CVPR2017	0.637	0.422
DaSiamRPN [21]	ECCV2018	0.735	0.512
UDT+ [15]	CVPR2019	0.650	0.457
MCCT [50]	CVPR2019	0.725	0.484
TADT [60]	CVPR2019	0.690	0.460
ASRCF [24]	CVPR2019	0.696	0.468
KAOT [67]	TMM2020	0.692	0.469
ARTracker [33]	GRSL2022	0.752	0.588
ALCF	Ours	0.694	0.480
DeepALCF	Ours	0.727	0.530

shows the comparison results. DeepALCF outperformed the third-best tracker DaSiamRPN [21] (0.512) by 1.8% in terms of success, and DeepALCF (0.727) is very close to DaSiamRPN (0.735) in terms of precision. Compared with ARTracker [33], ALCF reached a real-time speed of 51.3 FPS, which is far more than ARTracker (FPS: 13.28).

3.4.4. Results on UAV123@10fps

We conducted a comprehensive experiment in

Table 6. Overall evaluation of ALCF and some other state-of-the-art trackers on UAV123@10fps [40] (top 3 in bold).

Trackers	Precision	Success	Trackers	Precision	Success
HCF	0.601	0.426	MDNet	0.664	0.477
CFNet	0.568	0.422	IBCCF	0.651	0.481
SiamFC [68]	0.678	0.472	CREST	0.600	0.445
C-COT [69]	0.704	0.502	DeepSTRCF	0.680	0.499
DaSiamRPN	0.689	0.481	UDT+	0.674	0.470
MCCT	0.689	0.496	ADNet	0.647	0.422
TADT	0.685	0.507	ASRCF	0.685	0.477
ALCF	0.689	0.492	DeepALCF	0.712	0.516

, which compares DeepALCF and ALCF with numerous recent deep trackers. DeepALCF achieved the best precision (0.712) and success rate (0.516) scores. It can be found that ALCF with hand-drafted features obtained a third-best precision score, which outperformed a lot of deep trackers while running at over 50 FPS on a single CPU.

4. Conclusions

In this work, we proposed a novel tracking method called ALCF to deal with challenging situations (occlusion, deformation, etc.) in UAV tracking. Specifically, we introduced a novel criterion named target state estimation (TSE) to identify exact target states and adjust the learning rate for model update adaptively. In addition, a new model constraining regularization terms and an efficient lost-and-found strategy were further developed to enhance the robustness of the tracker to target variation and loss. Comprehensive experiments on four widely-used UAV datasets demonstrated that ALCF performs better than most state-of-the-art trackers in both precision and success scores while running over 50 FPS on a single CPU. We also provided a deep version tracker called DeepALCF, which is competitive against recently popularized deep trackers. We believe our work would contribute to the field of UAV tracking. We will focus on improving the real-time performance of the model in edge computing and running our algorithm on a hardware system on a UAV to evaluate its performance.