A Novel Ego-Motion Compensation Strategy for Automatic Target Tracking in FLIR Video Sequences taken from UAVs

Tracking targets in forward-looking infrared (FLIR) video sequences taken from airborne platforms is a challenging task. Several tracking failure modes can occur; in particular, discontinuities due to platform's motion can produce the so called ego-motion failure leading to unrecoverable errors in tracking the target. A novel ego-motion compensation technique for UAVs (unmanned aerial vehicles) is proposed. Data received from the autopilot can be used to predict the motion of the platform, thus allowing to identify a smaller region of the image (subframe) where the candidate target has to be searched for in the next frame of the sequence. The presented methodology is compared with a recently robust algorithm for automatic target tracking; experimental results show that the proposed motion estimation approach helps to improve performance both in terms of frames processed per second (targets are searched in smaller regions) and in terms of robustness (targets are correctly tracked for all the sequence's frames).     

EM-ML algorithm for track initialization using possibly noninformative data

Initializing and maintaining a track for a low observable (LO) (low SNR, low target detection probability and high false alarm rate) target can be very challenging because of the low information content of measurements. In addition, in some scenarios, target-originated measurements might not be present in many consecutive scans because of mispointing, target maneuvers, or erroneous preprocessing. That is, one might have a set of noninformative scans that could result in poor track initialization and maintenance. In this paper an algorithm based on the expectation-maximization (EM) algorithm combined with maximum likelihood (ML) estimation is presented for tracking slowly maneuvering targets in heavy clutter and possibly noninformative scans. The adaptive sliding-window EM-ML approach, which operates in batch mode, tries to reject or weight down noninformative scans using the Q-function in the M-step of the EM algorithm. It is shown that target features in the form of, for example, amplitude information (AI), can also be used to improve the estimates. In addition, performance bounds based on the supplemented EM (SEM) technique are also presented. The effectiveness of new algorithm is first demonstrated on a 78-frame long wave infrared (LWIR) data sequence consisting of an Fl Mirage fighter jet in heavy clutter. Previously, this scenario has been used as a benchmark for evaluating the performance of other track initialization algorithms. The new EM-ML estimator confirms the track by frame 20 while the ML-PDA (maximum likelihood estimator combined with probabilistic data association) algorithm, the IMM-MHT (interacting multiple model estimator combined with multiple hypothesis tracking) and the EVIM-PDA estimator previously required 28, 38, and 39 frames, respectively. The benefits of the new algorithm in terms of accuracy, early detection, and computational load are illustrated using simulated scenarios as well.     

Efficient target tracking using dynamic programming

A dynamic programming (DP) algorithm has been developed for the detection and tracking of subpixel-sized, low signal-to-noise ratio (SNR) targets observed by side-or forward-looking imaging sensors. A distinguishing feature of this approach is that target detection and tracking are combined into a single optimization procedure that takes into account statistical models of target motion, background noise, and clutter. Current work has led to a number of technical innovations that improve the performance and efficiency of the DP tracking algorithm, including the development of a new track scoring function, and an extension to the basic DP algorithm that reduces computation requirements by over an order of magnitude. A prototype infrared (IR) target tracking system incorporating these enhancements has been implemented for a step-starting IR camera application. Sensitivity improvements of several decibels over conventional sequential detection and tracking algorithms were realized     

Analysis of multiframe target detection using pixel statistics

Current track-before-detect (TBD) algorithms are developed and analyzed using a path statistic for each potential object trajectory. However this path statistic does not characterize overall performance gain. We propose a pixel-based statistic. This allows the TBD approach to be characterized as an image enhancement algorithm with detection gains compared with single frame detections. It is shown that for the TBD approach to have superior detection over single frame detection the target signal-to-noise ratio (SNR) must be greater than a threshold SNR in order to overcome the uncertainty in the target path. Tradeoffs are made for a class of velocity constrained target paths in terms of the detection gain with respect to the maximum target velocity and number of frames integrated     

Directed Subspace Search ML-PDA with Application to Active Sonar Tracking

The maximum likelihood probabilistic data association (ML-PDA) tracking algorithm is effective in tracking Very Low Observable targets (i.e., very low signal-to-noise ratio (SNR) targets in a high false alarm environment). However, the computational complexity associated with obtaining the track estimate in many cases has precluded its use in real-time scenarios. Previous ML-PDA implementations used a multi-pass grid (MPG) search to find the track estimate. Two alternate methods for finding the track estimate are presented-a genetic search and a newly developed directed subspace (DSS) search algorithm. Each algorithm is tested using active sonar scenarios in which an autonomous underwater vehicle searches for and tracks a target. Within each scenario, the problem parameters are varied to illustrate the relative performance of each search technique. Both the DSS search and the genetic algorithm are shown to be an order of magnitude more computationally efficient than the MPG search, making possible real-time implementation. In addition, the DSS search is shown to be the most effective technique at tracking a target at the lowest SNR levels-reliable tracking down to 5 dB (postprocessing SNR in a resolution cell) using a 5-frame sliding window is demonstrated, this being 6 dB better than the MPG search.     

Evaluating a multiple-hypothesis multitarget tracking algorithm

The performance evaluation of multiple-hypothesis, multitarget tracking algorithm is presented. We are primarily interested in target-detection/track-initiation capabilities as measures of performance. Through Monte Carlo simulations, a multiple-hypothesis tracking algorithm was evaluated in terms of 1) probability of establishing a track from target returns and 2) false track density. A radar was chosen as the sensor, and a general multiple-hypothesis, multitarget tracking algorithm was used in the Monte Carlo simulations. The simulation results predict the probability of establishing a track from returns of a target as well as the false track density per scan volume per unit time. The effects of the target radar cross section and the radar power, measured through the mean signal-to-noise ratio (SNR) were studied, as were the effects of detection threshold and track quality threshold. Computational requirements were also investigated     

自适应强杂波抑制与点状动目标检测

研究了基于自适应图像杂波抑制的微弱点状动目标检测技术。首先利用四叉树算法,将原始的非平稳图像分割成多个准平稳的图像子块,然后对各子块进行LS自适应背景杂波估计与抑制,从而获得准高斯白噪声背景;再利用目标运动连续性假设,将目标在相邻多帧上的位置状态模型化为高阶马尔可夫数据链,建立轨迹状态空间;根据该模型采用多帧沿轨迹非线性集成算法进行检测。既克服了传统的三维匹配算法造成搜索次数巨大的弱点,同时也避免了二维投影检测带来的信噪比下降。理论分析和大量仿真实验证明了其有效性。     

光电经纬仪红外成像跟踪处理器的研究与实现

以高性能数字信号处理器TMS320C6203为核心,结合CPLD（可编程逻辑器件）进行逻辑控制,用FPGA（现场可编程门阵列）进行预处理,设计了实时目标检测和跟踪系统。介绍了实时图像处理系统的硬件组成、工作原理、软件流程,重点分析目标检测算法和系统实时性。该系统被成功用于光电经纬仪红外图像处理系统中,经试验证明,系统对弱小目标的检测、识别和跟踪能力达到实际工程的实时性需求,大大提高了数据采集能力与处理速度,采样精度得到很大提高,完全满足系统设计要求。     

复杂环境下一种基于改进核相关滤波的视觉鲁棒目标跟踪方法

目标跟踪在自动驾驶和智能监控系统等实时视觉应用中发挥着重要作用。在遮挡、相似干扰等情况下，传统的基于相关滤波的跟踪算法容易发生漂移，鲁棒性有待进一步提高。基于此，提出了一种扩展特征描述的检测辅助核相关滤波目标跟踪架构。首先，在传统的核相关滤波目标跟踪算法的基础上，通过目标检测辅助对跟踪结果进行质量判断，实现对遮挡以及目标丢失的判别；然后通过拓展特征模板的构建与匹配，实现抗干扰相似目标判断及目标重定位；最终，以行人跟踪为例进行了试验，分别通过OTB数据及验证试验和移动机器人平台视觉跟踪验证试验，验证了算法的可行性，并对算法的跟踪性能进行了测试。试验结果表明，所提方法能够稳定地跟踪移动目标，对遮挡、相似干扰具有较强的鲁棒性。     

Passive ranging of a low observable ballistic missile in agravitational field

In this paper we present an estimation algorithm for tracking the motion of a low-observable target in a gravitational field, for example, an incoming ballistic missile (BM), using angle-only measurements. The measurements, which are obtained from a single stationary sensor, are available only for a short time. Also, the low target detection probability and high false alarm density present a difficult low-observable environment. The algorithm uses the probabilistic data association (PDA) algorithm in conjunction with maximum likelihood (ML) estimation to handle the false alarms and the less-than-unity target detection probability. The Cramer-Rao lower bound (CRLB) in clutter, which quantifies the best achievable estimator accuracy for this problem in the presence of false alarms and nonunity detection probability, is also presented. The proposed estimator is shown to be efficient, that is, it meets the CRLB, even for low-observable fluctuating targets with 6 dB average signal-to-noise ratio (SNR). For a BM in free flight with 0.6 single-scan detection probability, one can achieve a track detection probability of 0.99 with a negligible probability of false track acceptance     

空间多分辨率模糊目标跟踪

提出了一种新的模糊目标跟踪算法--CPDA算法。这个算法在空间多分辨率框架下应用概率数据互联算法,在粗分辨率上实现模糊目标跟踪。在不同虚警密度的模糊目标环境下,利用仿真实验分析了CPDA算法的跟踪性能,同时将其与单分辨率上的联合概率数据互联方法进行了性能比较。仿真结果表明,CPDA算法的跟踪性能在达到与单分辨率上JPDA算法同样性能的条件下,能够以较小的计算量跟踪模糊目标。     

基于邻帧差分近邻反相特征的红外运动点目标检测算法

基于运动点目标在邻帧差分图像中所具有的近邻反相特征,即运动点目标的两个位置相邻近、灰度值一正一负,提出一种在复杂背景下,基于红外序列图像的运动点目标检测算法.本算法利用该特征在邻帧差分图像中检测反相点对,进而构造反相点对矢量图,最后依据累积反相点对矢量图中多矢量首位相接的连续性检测出运动的点目标.文中给出并证明应用本算法能以概率1检测到运动点目标的收敛性定理.对典型复杂背景下10幅1000帧图像的仿真结果表明,当信噪比大于或等于1.5时,可以有效检测出运动点目标.     

基于注意力机制和特征匹配的空对地多目标检测与跟踪算法

多目标跟踪算法是实现无人机自主导航的关键技术，为解决现有方法存在的小目标检测能力弱、计算能耗大、鲁棒性差等问题，提出一种基于注意力机制和特征匹配的多目标空对地跟踪算法，以实现航拍视角下对目标的精准高效跟踪。首先，引入通道可分离卷积，实现目标检测模型的轻量化；其次，构造融合空间注意力机制的小目标检测分支，提高对小微目标的检测精度，最后，优化目标跟踪算法的外观重识别网络，提高多目标跟踪效率。使用Visdrone2019-MOT数据集对所提算法进行验证，实验结果表明，所提算法的MOTA值提高了0.６%，FPS值为21.31帧/s，在模型大小和跟踪精度上实现了较好的平衡。     

Dynamic Programming Solution for Detecting Dim Moving Targets Part II: Analysis

An on-board mosaic sensor is staring down from a satellite to afixed point on the ground while collecting frames that contain targetsignatures and background noise. A dynamic programmingalgorithm (DPA) has been developed to optimally detect dim movingtargets that cross the sensor's field of view. The algorithm has beendescribed in Part I of this paper. Here, in Part II, we analyze thealgorithm performance and compare it with that of a simpleassembling of thresholded frames; the DPA's SNR performance isshown to be at least an order of magnitude better.     

Track-before-detect procedures for early detection of moving target from airborne radars

In this paper we present a family of track-before-detect (TBD) procedures for early detection of moving targets from airborne radars. Upon a sectorization of the coverage area, the received echoes are jointly processed in the azimuth-range-Doppler domain and in the time domain through a Viterbi-like algorithm that exploits the physically admissible target transitions between successive illuminations, in order to collect all of the energy back-scattered during the time on target (TOT). A reduced-complexity implementation is derived assuming, at the design stage, that the target does not change resolution cell during the TOT in each scan. The constant false alarm rate (CFAR) constraint is also englobed in the proposed procedures as well as the possibility of working with quantized data. Simulation results show that the proposed algorithms have good detection and tracking capabilities even for high target velocities and low quantization rates.     

Maximum likelihood track-before-detect with fluctuating targetamplitude

An important problem in target tracking is the detection and tracking of targets in very low signal-to-noise ratio (SNR) environments. In the past, several approaches have been used, including maximum likelihood. The major novelty of this work is the incorporation of a model for fluctuating target amplitude into the maximum likelihood approach for tracking of constant velocity targets. Coupled with a realistic sensor model, this allows the exploitation of signal correlation between resolution cells in the same frame, and also from one frame to the next. The fluctuating amplitude model is a first order model to reflect the inter-frame correlation. The amplitude estimates are obtained using a Kalman filter, from which the likelihood function is derived. A numerical maximization technique avoids problems previously encountered in “velocity filtering” approaches due to mismatch between assumed and actual target velocity, at the cost of additional computation. The Cramer-Rao lower bound (CRLB) is derived for a constant, known amplitude case. Estimation errors are close to this CRLB even when the amplitude is unknown. Results show track detection performance for unknown signal amplitude is nearly the same as that obtained when the correct signal model is used     

Integrated track splitting filter--efficient multi-scan single target tracking in clutter

A fully automatic tracking algorithm must be able to deal with an unknown number of targets, unknown target initiation and termination times, false measurements and possibly time-varying target trajectory behaviour. An efficient algorithm for tracking in this environment is presented here. This approach makes use of estimates of the probability of target existence, which is an integral part of the algorithm. This allows for the efficient generation and management of possible target hypotheses, yielding an algorithm with performance that matches what can be obtained by multiple hypothesis tracking-based approaches, but at a significantly lower computational cost. This paper considers only the single target case for clarity. The extension to multiple targets is easily incorporated into this framework. Simulation studies are given that show the effectiveness of this approach in the presence of heavy and nonuniform clutter when tracking a target in an environment of low probability of detection and in an environment where the target performs violent manoeuvres.     

Bearings-only tracking of maneuvering targets using abatch-recursive estimator

We present a new batch-recursive estimator for tracking maneuvering targets from bearings-only measurements in clutter (i.e., for low signal-to-noise ratio (SNR) targets), Standard recursive estimators like the extended Kalman Iter (EKF) suffer from poor convergence and erratic behavior due to the lack of initial target range information, On the other hand, batch estimators cannot handle target maneuvers. In order to rectify these shortcomings, we combine the batch maximum likelihood-probabilistic data association (ML-PDA) estimator with the recursive interacting multiple model (IMM) estimator with probabilistic data association (PDA) to result in better track initialization as well as track maintenance results in the presence of clutter. It is also demonstrated how the batch-recursive estimator can be used for adaptive decisions for ownship maneuvers based on the target state estimation to enhance the target observability. The tracking algorithm is shown to be effective for targets with 8 dB SNR     

SMALL TARGET TRACKING TECHNIQUE WITH DATA FUSION OF DISTRIBUTED SENSOR NET

ＳＭＡＬＬＴＡＲＧＥＴＴＲＡＣＫＩＮＧＴＥＣＨＮＩＱＵＥＷＩＴＨＤＡＴＡＦＵＳＩＯＮＯＦＤＩＳＴＲＩＢＵＴＥＤＳＥＮＳＯＲＮＥＴＣＨＥＮＧＨｏｎｇｗｅｉ（程洪玮）,ＺＨＯＵＹｉｙｕ（周一宇）,ＳＵＮＺｈｏｎｇｋａｎｇ（孙仲康）（Ｆａｃｕｌｔｙ４０６,...     

Sequential track extraction

Sensors like radar or sonar usually produce data on the basis of a single frame of observation: target detections. The detection performance is described by quantities like detection probability Pd and false alarm density f. A different task of detection is formation of tracks of targets unknown in number from data of multiple consecutive frames of observation. This leads to quantities which are of a higher level of abstraction: extracted tracks. This again is a detection process. Under benign conditions (high Pd, low f and well separated targets) conventional methods of track initiation are recommended to solve a simple task. However, under hard conditions the process of track extraction is known to be difficult. We here concentrate on the case of well separated targets and derive an optimal combinatorial method which can be used under hard operating conditions. The method relates to MHT (multiple hypothesis tracking), uses a sequential likelihood ratio test and derives benefit from processing signal strength information. The performance of the track extraction method is described by parameters such as detection probability and false detection rate on track level, while Pd and f are input parameters which relate to the signal-to-noise interference ratio (SNIR), the clutter density, and the threshold set for target detection. In particular the average test lengths are analyzed parametrically as they are relevant for a user to estimate the time delay for track formation under hard conditions