Thresholding and energy-based techniques are not suitable in the targetdetection task because the presence foreign objects and non-controlled light conditions affect the result. On the other hand, machine learning techniques need to be trained taking into account all (if possible) image degradations and the number of the landing targets in the set to be successful. In this work, we performed the landing targetdetection and recognition by way of a non-supervised perceptual model. The landing target model is built using the Gestalt’s visual perception principles under the idea that the contours characteristics behave as outliers in the multi-feature distribution. The advantages of this method lie in the versatility of features that can be used to carry out the "groupings" without the need to establish regulation parameters or to have a training phase. Besides, the proposed method is robust to changes in lighting, the presence of shadows and noise in the image and has the auto-correction capacity.
4 Department of Physics and Astronomy, School of Natural Sciences, University of Manchester, Manchester M13 9PL, United Kingdom 5 Department of Physics, University of Ottawa, Ottawa, Ontario, Canada K1N 6N5
(Received 16 January 2020; accepted 6 April 2020; published 4 May 2020)
In this work we investigate quantum-enhanced targetdetection in the presence of large background noise using multidimensional quantum correlations between photon pairs generated through spontaneous parametric down- conversion. Until now similar experiments have only utilized one of the photon pairs’ many degrees of freedom such as temporal correlations and photon number correlations. Here, we utilized both temporal and spectral correlations of the photon pairs and achieved over an order of magnitude reduction to the background noise and in turn significant reduction to data acquisition time when compared to utilizing only temporal modes. We believe this work represents an important step in realizing a practical, real-time quantum-enhanced targetdetection system. The demonstrated technique will also be of importance in many other quantum sensing applications and quantum communications.
4 Conclusion and Future Extensions
We have described the procedure for the landing targetdetection and recog- nition based on a perception model. The algorithm is based on the Helmholtz non-accidentalness principle and the Gestalt theory. The non-accidentalness esti- mation is performed in a multi-feature object space built from the image contours at different scales. This approach allows us to obtain scene information avoiding the loss of information because of the objects’ change of size or the presence of shadows and noise. We have used the similarity and proximity Gestalt laws to group the contours and build a perceptual object and the Hamming error codes to perform the landing target recognition. The experiments show that the proposed methodology for the detection of landing targets is robust to un- controlled light conditions and other images degradations existing in complex environments.
One-step Generalized Likelihood Ratio Test for Subpixel TargetDetection in Hyperspectral Imaging
Franc¸ois Vincent and Olivier Besson
Abstract—One of the main objectives of hyperspectral image processing is to detect a given target among an unknown background. The standard data to conduct such a detection is a reflectance map, where the spectral signatures of each pixel’s components, known as endmembers, are associated with their abundances in the pixel. Due to the low spatial resolution of most hyperspectral sensors, such a target occupies a fraction of the pixel. A widely used model in case of subpixel targets is the replacement model. Among the vast number of possible detectors, algorithms matched to the replacement model are quite rare. One of the few examples is the Finite Target Matched Filter, which is an adjustment of the well-known Matched Filter. In this paper, we derive the exact Generalized Likelihood Ratio Test for this model. This new detector can be used both with a local covariance estimation window or a global one. It is shown to outperform the standard target detectors on real data, especially for small covariance estimation windows.
6 CONCLUSION AND FUTURE WORK
To the best of our knowledge, this paper presents one of the first POMDP-based implementations of targetdetection and recognition mission by an autonomous rotorcraft UAV. Our contribution is three- fold: (i) we model perception and mission actions in the same de- cision formalism using a single POMDP model; (ii) we statistically learn a meaningful probabilistic observation model of the outputs of an image processing algorithm that feeds the POMDP model; (iii) we provide practical algorithmic means to optimize and exe- cute POMDP policies in parallel under time constraints, what is re- quired because the POMDP problem is generated during the flight. We analyzed experiments conducted with a realistic “hardware in the loop” simulation based on real data, which demonstrate that POMDP planning techniques are now mature enough to tackle complex aerial robotics missions, assuming the use of some kind of “optimize- while-execute” framework, as the one proposed in this paper.
Indeed, our aim was to control for the fundamental process of perceiving rhythm and to focus, instead, on the effects of variations in beat saliency within a given rhythm on accompanying behavior. According to the DAT ( Jones), stimuli presented at strong versus weak beats differ in terms of the amount of attentional energy allocated toward them. Strong beats capture attention, thus favoring the processing of any stimuli that happen to occur at those moments in time. Two recent fMRI studies have reported that the behavioral facilitation in- duced by rhythmic stimulus presentation was accom- panied by increased activity in BG (Geiser et al., 2012) as well as bilateral pFC, insula, and left IPC (Marchant & Driver, 2012). However, in both of these studies, rhythmic sequences were compared with nonrhythmic ones, there- by confounding basic perception of rhythm with the use of this rhythm to optimize behavior. In our study, we de- liberately controlled for rhythm perception processes, thereby isolating activity related to beat-induced variations in the allocation of attention to precise moments in time more selectively. Our results indicate that, whereas BG and SMA may be activated by the perception of rhythm in the first place (e.g., Grahn & Brett, 2007), the left IPC is activated whenever temporally salient elements of that rhythm capture attention, thereby optimizing the process- ing of stimuli occurring at that time. This neuroanatomical distinction is an example of the functional difference be- tween timing to estimate duration (perceiving rhythmicity) and timing to optimize sensorimotor processing (using the temporal predictability of the rhythm to improve targetdetection; Coull & Nobre, 2008).
This paper tackles high-level decision-making tech- niques for robotic missions, which involve both ac- tive sensing and symbolic goal reaching, under uncer- tain probabilistic environments and strong time con- straints. Our case study is a POMDP model of an on- line multi-targetdetection and recognition mission by an autonomous UAV. The POMDP model of the multi- targetdetection and recognition problem is generated online from a list of areas of interest, which are auto- matically extracted at the beginning of the flight from a coarse-grained high altitude observation of the scene. The POMDP observation model relies on a statistical abstraction of an image processing algorithm’s output used to detect targets. As the POMDP problem cannot be known and thus optimized before the beginning of the flight, our main contribution is an “optimize-while- execute” algorithmic framework: it drives a POMDP sub-planner to optimize and execute the POMDP pol- icy in parallel under action duration constraints. We present new results from real outdoor flights and SAIL simulations, which highlight both the benefits of using POMDPs in multi-targetdetection and recognition mis- sions, and of our “optimize-while-execute” paradigm.
are used, on the rationale that the background there is more representative while the background in farther pixels may differ from that in PUT [3, 4, 5].
Concerning the choice of the background distribution, the Gaussian assumption prevails, probably due to the huge amount of methods that have been developed for previous applications, such as radar, and its mathematical tractability that enables straightforward derivations and analytical performance evaluations. Thereby, many targetdetection schemes can be used in this context, such as the adaptive matched filter (AMF)  or Kelly’s detector , to name a few. They correspond to two different approaches to derive the generalized likelihood ratio test (GLRT). Kelly’s detector is known as a one-step GLRT, as it has been derived directly from the joint distribution of both the PUT and the training samples, whereas the AMF is its two-step counterpart, namely derived from the PUT distribution, assuming that the background parameters are known and then replaced by their estimates from the training samples.
(a) Input image (b) Filtering (c) Car detection (d) Matching
Figure 1: Targetdetection and recognition image processing based on (Saux and Sanfourche 2011).
in (Teichteil-Konigsbuch, Lesire, and Infantes 2011), previ- ously restricted to deterministic or MDP planning, to on-line solve large POMDPs under time constraints. Our extension is a meta planner that relies on standard POMDP planners like PBVI, HSVI, PERSEUS, AEMS, etc., which are called from possible future execution states while executing the current optimized action in the current execution state, in anticipation of the probabilistic evolution of the system and its environment. One of the issues of our extension was to adapt the mechanisms of (Teichteil-Konigsbuch, Lesire, and Infantes 2011) based on completely observable states, to be- lief states and point-based paradigms used by many state-of- the-art POMDP planners (Pineau, Gordon, and Thrun 2003; Ross and Chaib-Draa 2007). This framework is differ- ent from real-time algorithms like RTDP-bel (Bonet and Geffner 2009) that solve the POMDP only from the current execution state, but not from future possible ones as we pro- pose.
I N THE last few decades, spurred by analysis of experi- mental radar data , , , , , considerable atten- tion has been focused on targetdetection in heavy-tailed clutter environment, see e.g., , , ,  and references therein. Most studies deal with detection of non-ﬂuctuating over co- herent integration time (CIT) target return in a clutter that can be presented as a compound-Gaussian process ,  or, more generally, is assumed to follow an elliptically contoured distribution . Traditionally the detection problem is formu- lated from a single resolution cell. Herein, we consider long observation interval (“stare mode”) that signiﬁcantly exceeds the target and clutter ﬂuctuations spectra Nyquist rates, which can be recast as the following hypotheses testing problem (
B. Target Tracking
The second major class of problems, target tracking, corre- sponds to the tasks that arise when one or several targets have been detected or assigned – often following the success of targetdetection tasks. Here, coping with a target may imply keeping it in sight, to provide information on it (mainly to localise it over time, but identifying it can also be an objective), or to catch it. In all cases tracker robots need to stay “close” to the targets, the required distance being zero when it comes to catch the targets. Coping with a single target may require one or more robots, depending on the context. It is for instance preferable to have multiple vantage points on each target to refine their locations. We refer to this latter class of problems as target localization problems. We also distinguish the one vs. one problems (following) from the multi-robot multi-target problems (observation).
a b s t r a c t
One of the main issue in detecting a target from an hyperspectral image relies on properly identifying the background. Many assumptions about its distribution can be advocated, even if the Gaussian hypothesis prevails. Nevertheless, the huge majority of the resulting detection schemes assume that the background distribution remains the same whether the target is present or not. In practice, because of the spectral variability of the target and the non-linear mixing with the background radiance, this hypothesis is not strictly true. In this paper, we consider that an unknown background mismatch exists between the two hypotheses. Under the assumption that this mismatch is small, we derive an approximation of the Like- lihood Ratio for the problem at hand. This general formulation is then applied to the case of Gaussian distributed background, leading to a robust Adaptive Matched Filter. The behaviour of this new detec- tor is analysed and compared to popular detectors. Numerical simulations, based on real data, show the possible improvement in case of target signature mismatch.
*** University of Rennes1, France
Abstract: The goal of most modern automotive safety driver assistance functions is to avoid possible collisions. Pedestrian protection, predictive emergency braking or turn and crossing assist functions are usually based on two steps. First, the radar provides detailed information on the environment, and then a detection procedure is driven. Because of the complicated environment near the vehicle, this second step is a difficult task to achieve in order to give reliable information to the driver. In this paper, we propose to fuse the environment estimation and the detection step into a simple and direct collision target detector. Indeed, this procedure allows detecting possible collision targets, based on their typical Doppler signature, while rejecting all fixed and non-dangerous targets (clutter). Moreover, this detector only exploits a single antenna, and the necessary target speed vector information is obtained by a second order phase expansion using a long integration time, making the use of an antenna array unnecessary.
is being optimised. Whenever possible, feeds through this shield should be designed to avoid direct sight.
3.3 Target Heat Exchanger and Upper Liquid Metal Container
The Target Heat Exchanger (THE) is made of 12 pins of 120 cm long arranged in circle. Each pin consists of 2 tubes concentrically arranged: the inlet and outlet tubes. Using the diathermic oil Diphyl THT ® as cooling medium (cf. section 5.4.3 LBE-organic oil interaction), it was necessary to implement a spiral in the oil path to increase the contact length. The main problem in the design of the THE was to comply with the complex thermal conditions and to limit the resulting thermo- mechanical stresses. A good agreement has been reached by connecting the pins to the inlet and outlet oil distribution boxes by flexible bellows and by inserting thin shrouds as heat shields. The heat is removed from the THE by an intermediate oil loop designed by Ansaldo. An intermediate water-cooling loop designed and built by PSI then evacuates the heat from the oil loop. The major components of THX are (see figure 3-4):
Ligation probe design. For the LDR, we designed specific probes for the 16S
rRNA gene sequences of 19 different cyanobacterial groups. These groups were identified by using a cyanobacterial 16S rRNA gene alignment built with ARB software, version Beta 011107 (16). The alignment contained 281 sequences from public databases and 57 from this study in addition to the out-group Escherichia coli. All of these sequences were longer than 1,400 bp, except the two sequences of Antarctic Phormidium (about 1,350 bp) and 21 (of 42) sequences of Prochlo- rococcus marinus (about 1,250 bp). All sequences were aligned with CLUSTAL W (26) and ARB. The sequence alignment is available upon request. The phylogenetic analysis was performed with ARB by using the neighbor-joining (NJ) algorithm (22). From the sequence alignment, group-specific consensus sequences were obtained with a cutoff percentage of 75%. If a base at a given position occurred at a lower frequency than the cutoff percentage, it was replaced by an appropriate International Union of Pure and Applied Chemistry ambiguity code in the consensus sequence. The group-specific consensus sequences were imported to GCG Omiga, version 2.0 (Oxford Molecular Ltd.), for group-specific probe design. The probes were designed by following the LDR approach. After hybridization of a discriminating probe and a common probe to the target sequence, ligation occurs only if there is perfect complementarity between the two probes and the template, in this case, an amplified fragment of the 16S rRNA gene (Fig. 1). For this reason, the discriminating probes were designed to have 3⬘ ends unique to each of the 19 cyanobacterial groups. The common probes were located immediately after the discriminating probes according to the group-specific consensus sequences. An example of selection is shown in Fig. 2. To discard potentially unspecific probe pairs, we checked each probe pair (dis- criminating probe and common probe) by using the probe match tool of the ARB program. We also designed a probe pair (named UNICYANO) to detect the presence of any cyanobacteria in the sample. No significant self-annealing of the probe sequences was detected by computer analysis (data not shown). All probes were designed to have a theoretical melting temperature (T m ) between 63 and
The highest ranked model for the whole gp37-gp38 multimer was TS086_1 by the BAKER group (Figure 7F) with a QS-score of 0.37. Despite failing to predict gp38 and its attachment to gp37 correctly, this model very accurately determined the composition of the gp37 β-helix, including the N-terminal triangular and C-terminal interdigi- tated domains. As a single target, gp37 (T0953s1-D1) was similarly predicted well, with the top two models by groups A7D and BAKER (GDT-TS of 54.48 and 48.88, respectively) shown in Figure 7D. Visual F I G U R E 7 The adhesin tip of the Salmonella phage S16 long tail fiber. A, Transmission electron micrograph of phage S16 with arrows (1) pointing to the approximate location of gp38 at the tip of the LTF and (2) pointing to the baseplate. B, Cartoon representation of the LTF distal tip complex of homotrimeric gp37 β-helix (cyan, magenta, pink) attached to a single gp38 adhesin (gray) with the structurally unique “polyglycine sandwich ” domain rainbow colored (blue to red). Gp38 connects to gp37 through hydrophobic interactions, involving three highly conserved tryptophan residues on the apex of each α-helix of the gp38 attachment domain (yellow sticks) that occupy three symmetry-equivalent hydrophobic pockets on the gp37 base. C, Head-on view of the polyglycine sandwich domain formed by the 10 glycine-rich motifs (GRMs) of gp38 folded into a three-layered lattice of PG II helices. Labeled are the five distal loops formed by HVSs that form the (yet unknown) receptor-
Figure 1.1: Aerosonde: First robotic aircraft to cross the North Atlantic.
some of them.
The second motivation comes from the nature of most UAV applications. They come in a widening range: airborne surveillance, military information gathering, military offensive actions, automated search and rescue, gas pipe line monitoring and automated forest fire surveillance. The common point to most of these applications is that they are defined with respect to the ground environment: ground target tracking, be the target static, slowly moving or maneuvering at high speeds, is an essential task for UAVs. For such tasks, fixed wing UAVs have the advantage over helicopters to reach high speeds at low energy consumption, but their dynamics constrain the visibility of the ground target. This thesis considers this issue, and introduces various strategies to achieve ground visual target tracking, aiming at maximizing the target visibility.
Though the CPHD filter’s model for birth targets has the potential to address spawning targets , there may be cases where specific spawning models are more applicable. Con- sider for example the case of tracking resident space objects (RSOs), natural and artificial Earth orbiting satellites consisting of active spacecraft, decommissioned payloads, and debris. In this context, spawning events include the deployment of Cube- Sats from a launch vehicle ,  and fragmentation events caused by the unintentional  or intentional  collision of objects. Without spawning, the best option may be the use of diffuse birth regions, however, the volume of space to be filled requires a potentially intractable number of birth regions . To improve the CPHD filter’s performance for space- object tracking, previous research has presented a measurement- based birth model that leverages an astrodynamics approach to track initialization for RSOs . While such an approach may be effective for tracking spawned RSOs, a multi-target fil- ter that more accurately describes the physical processes that produce new RSOs through a specific spawning model is ex- pected to provide better accuracy and faster confirmation of new objects.