We present a real-time multi-object tracker using an enhanced version of the Gaussian mixture probability hypothesis density (GM-PHD) filter to track detections of a state-of-the-art convolutional neural network (CNN). This approach adapts the GM-PHD filter to a real-world scenario to recover target trajectories in remote sensing videos. Our GM-PHD filter uses a measurement-driven birth, considers past tracked objects, and uses CNN information to propose better hypotheses initialization. Additionally, we present a label tracking solution for the GM-PHD filter to improve identity propagation given target path uncertainties. Our results show competitive scores against other trackers while obtaining real-time performance. Code is available at https://github.com/Ayana-Inria/RFS-filters-for-satellite-videos.

The SPOT 6-7 satellite ground segment includes a systematic and automatic cloud detection step in order to feed a catalogue with a binary cloud mask and an appropriate confidence measure. In order to significantly improve the SPOT cloud detection and get rid of frequent manual re-labelings, we study a new automatic cloud detection technique that is adapted to large datasets. The proposed method is based on a modified distributed boosting algorithm. Experiments conducted using the framework Apache Spark on a SPOT 6 image database with various landscapes and cloud coverage show promising results.

In this paper, we present a novel approach based on spatio-temporal marked point processes to detect and track moving objects in a batch of high resolution images. Batch processing techniques are applicable to and desirable for a large class of applications such as offline scene and video analysis, and provide better overall detection and data association accuracy than sequential methods. We develop a new, intuitive energy based model consisting of several terms that take into account both the image evidence and physical constraints such as target dynamics, track persistence and mutual exclusion. We construct a suitable optimization scheme that allows us to find strong local minima of the proposed highly non-convex energy. We test our model on three batches of 25 synthetic biological images with different levels of noise. Our main application however consists of two batches of 14 remotely sensed high resolution optical images of boats which are particularly hard to analyze due to the different angles at which the images were taken and the low temporal frequency.

Pushbroom cameras are widely used for earth observation applications. This sensor acquires 1D images over time and uses the straight motion of the satellite to sweep out a region of space and build a 2D image. The stability of the satellite is critical during the pushbroom acquisition process. Therefore its attitude is assumed to be constant overtime. However, the recent manufacture of smaller and lighter satellites to reduce launching cost has weakened this assumption. Small oscillations of the satellite's attitude can result in noticeable warps in images, and geolocation information is lost as the satellite does not capture what it ought to. Current solutions use inertial sensors to control the attitude and correct the images, but they are costly and of limited precision. As the warped images do contain information about attitude variations, we suggest using image registration to estimate them. We exploit the geometry of the focal plane and the stationary nature of the disturbances to recover undistorted images. We embed the estimation in a Bayesian framework where image registration, a prior on attitude variations and a radiometric correction model are fused to retrieve the motion of the satellite. We illustrate the performance of our algorithm on four satellite datasets.

Linear pushbroom cameras are widely used in passive remote sensing from space as they provide high resolution images. In earth observation applications, where several pushbroom sensors are mounted in a single focal plane, small dynamic disturbances of the satellite's orientation lead to noticeable geometrical distortions in the images. In this paper, we present a global method to estimate those disturbances, which are effectively vibrations. We exploit the geometry of the focal plane and the stationary nature of the disturbances to recover undistorted images. To do so, we embed the estimation process in a Bayesian framework. An autoregressive model is used as a prior on the vibrations. The problem can be seen as a global image registration task where multiple pushbroom images are registered to the same coordinate system, the registration parameters being the vibration coefficients. An alternating maximisation procedure is designed to obtain Maximum a Posteriori estimates (MAP) of the vibrations as well as of the autoregressive model coefficients. We illustrate the performance of our algorithm on various datasets of satellite imagery1.

Satellite image registration has been investigated for several years. Nevertheless, little attention has been paid to the linear geometry of the satellite's imaging sensor, often consisting of several pushbroom cameras. Each pushbroom camera captures 1-D image and uses straight motion of the satellite to build a 2-D image. Yet, attitude variations of the satellite during the aquisition process can lead to significant distortions in the 2-D image. In this paper, we expose the problem and present a constrained image registration method to estimate the satellite's attitude variations, and thus correct the distorted images. We use a Lucas Kanade framework and a piecewise polynomial model under constraints to deduce the registration equation. The performances of our algorithm are shown on two satellite datasets.

Pushbroom sensors are widely used for earth observation from space; this camera captures 1-D image and uses straight motion of the satellite to build a 2-D image. Yet, attitude variations of the satellite during the aquisition process can lead to significant distortions in the 2-D image. Current solutions use inertial sensors to record attitude variations and correct images, but they are expensive and not accurate enough for high frequency variations. In this paper, we expose the problem and present an image registration method to estimate the satellite's attitude variations, and thus correct the distorted images. We use a Lucas Kanade framework to deduce the registration equation; the performances of our algorithm are shown on two satellite datasets.

Simulation, detection and estimation of the spread of a biochemical substance are key elements in environmental monitoring. Solving these problems are important for efficient decontamination purposes and prediction of the cloud evolution. We present a set of tools describing the measurements of an array of biochemical sensors through a physical dispersion model, which is amenable to statistical analysis. We first approximate the dispersion model of a contaminant in a realistic environment (for instance urban) through numerical simulations of reflected stochastic diffusions describing the microscopic transport phenomena due to wind and chemical diffusion using the Feynmann-Kac formula. Second, we propose a Bayesian framework based on a random field for localizing multiple dispersive sources with small amounts of measurements. Third, we present a sequential detector allowing on-line analysis and detecting whether a change has occurred, based on realistic numerical simulation. Numerical examples illustrate our results for a dispersion among buildings.

This work presents a framework for automatic feature extraction from images using stochastic geometry. Features in images are modeled as realizations of a spatial point process of geometrical shapes. This framework allows the incorporation of a priori knowledge on the spatial repartition of features. More specifically, we present a model based on the superposition of a process of segments and a process of rectangles. The former is dedicated to the detection of linear networks of discontinuities, while the latter aims at segmenting homogeneous areas. An energy is defined, favoring connections of segments, alignments of rectangles, as well as a relevant interaction between both types of objects. The estimation is performed by minimizing the energy using a simulated annealing algorithm. The proposed model is applied to the analysis of Digital Elevation Models (DEMs). These images are raster data representing the altimetry of a dense urban area. We present results on real data provided by the IGN (French National Geographic Institute) consisting in low quality DEMs of various types.

In this work, we propose a new model based on stochastic geometry for extracting features from images. This type of model allows the incorporation of a prior knowledge on the interactions between features within the extraction process. We focus on the specific problem of automatic building extraction from digital elevation models (DEMs). The model we propose is based on two interacting spatial point processes, the former being a process of rectangles, the latter a process of segments. An energy associated with the resulting process is defined. This energy consists in five main parts. We first define two energy data terms to make the rectangles fit the homogeneous areas and the segments fit meaningful discontinuities. Two prior terms favoring respectively the alignment of rectangles and the connection of segments are incorporated. The last part of the energy is an interaction term that makes the two types of objects cooperate. We present results on real data provided by the IGN (French Geographic Institute)

To extract buildings from digital elevation models, we define a point process whose points represent buildings. We then define a density for this point process which is split into two parts, consisting in an "internal field" that allows us to model the prior knowledge we have on patterns of buildings in urban areas and an "external field" that makes the point process fit the data. We then use a metropolis Hastings Green sampler coupled with a simulated annealing that gives the configuration of buildings minimizing the energy we have defined. We present results on real data provided by the French Mapping Institute (IGN).