Daniel A. Lavigne

Robust vehicle detection in low-resolution aerial imagery

We propose a feature-based approach for vehicle detection in aerial imagery with 11.2 cm/pixel resolution. The approach is free of all constraints related to the vehicles appearance. The scale-invariant feature transform (SIFT) is used to extract keypoints in the image. The local structure in the neighbouring of the SIFT keypoints is described by 128 gradient orientation based features. A Support Vector Machine is used to create a model which is able to predict if the SIFT keypoints belong to or not to car structures in the image. The collection of SIFT keypoints with car label are clustered in the geometric space into subsets and each subset is associated to one car. This clustering is based on the Affinity Propagation algorithm modified to take into account specific spatial constraint related to geometry of cars at the given resolution.

Target discrimination of man-made objects using passive polarimetric signatures acquired in the visible and infrared spectral bands

Surveillance operations and search and rescue missions regularly exploit electro-optic imaging systems to detect targets of interest in both the civilian and military communities. By incorporating the polarization of light as supplementary information to such electro-optic imaging systems, it is possible to increase their target discrimination capabilities, considering that man-made objects are known to depolarized light in different manner than natural backgrounds. As it is known that electro-magnetic radiation emitted and reflected from a smooth surface observed near a grazing angle becomes partially polarized in the visible and infrared wavelength bands, additional information about the shape, roughness, shading, and surface temperatures of difficult targets can be extracted by processing effectively such reflected/emitted polarized signatures. This paper presents a set of polarimetric image processing algorithms devised to extract meaningful information from a broad range of man-made objects. Passive polarimetric signatures are acquired in the visible, shortwave infrared, midwave infrared, and longwave infrared bands using a fully automated imaging system developed at DRDC Valcartier. A fusion algorithm is used to enable the discrimination of some objects lying in shadowed areas. Performance metrics, derived from the computed Stokes parameters, characterize the degree of polarization of man-made objects. Field experiments conducted during winter and summer time demonstrate: 1) the utility of the imaging system to collect polarized signatures of different objects in the visible and infrared spectral bands, and 2) the enhanced performance of target discrimination and fusion algorithms to exploit the polarized signatures of man-made objects against cluttered backgrounds.

A new passive polarimetric imaging system collecting polarization signatures in the visible and infrared bands

Electro-optical imaging systems are frequently employed during surveillance operations and search and rescue missions to detect various targets of interest in both the civilian and military communities. By incorporating the polarization of light as supplementary information to such electro-optical imaging systems, it may be possible to increase the target discrimination performance considering that man-made objects are known to depolarize light in different manners than natural backgrounds. Consequently, many passive Stokes-vector imagers have been developed over the years. These sensors generally operate using one single spectral band at a time, which limits considerably the polarization information collected across a scene over a predefined specific spectral range. In order to improve the understanding of the phenomena that arise in polarimetric signatures of man-made targets, a new passive polarimetric imaging system was developed at Defence Research and Development Canada - Valcartier to collect polarization signatures over an extended spectral coverage. The Visible Infrared Passive Spectral Polarimetric Imager for Contrast Enhancement (VIP SPICE) operates four broad-band cameras concomitantly in the visible (VIS), the shortwave infrared (SWIR), the midwave infrared (MWIR), and the longwave infrared (LWIR) bands. The sensor is made of four synchronously-rotating polarizers mounted in front of each of the four cameras. Polarimetric signatures of man-made objects were acquired at various polarization angles in the four spectral bands. Preliminary results demonstrate the utility of the sensor to collect significant polarimetric signatures to discriminate man-made objects from their background.

Enhanced Military Target Discrimination using Active and Passive Polarimetric Imagery

Surveillance operations often make use of electro-optic (EO) imaging systems to detect civilian and military targets. To increase the overall target detection performance, such active/passive EO sensors could exploit the polarization of light as additional information to discriminate man made objects against different backgrounds. The target contrast enhancement obtained by analyzing the polarization of the reflected light from either a direct polarized laser source as encountered in active imagers, or from natural ambient illumination, can be used for such target discrimination scheme. This paper reports results from field experiments exploiting polarization-based imaging sensors to enhance the detection of man made objects. Active and passive polarimetric signatures of objects have been acquired at wavelengths in the near and long-wave infrared bands. Results demonstrate to what extent and under which illumination and environmental conditions the exploitation of active/passive polarimetric images is suitable to enable target discrimination.

Evaluation of active and passive polarimetric electro-optic imagery for civilian and military targets discrimination

Electro-optic (EO) imaging systems are commonly used to detect civilian and military targets during surveillance operations and search and rescue missions. Adding the polarization of light as additional information to such active and passive EO imaging systems may increase the target discrimination performance, as man made objects are known to depolarized light in different manner than natural background. However, while the polarization of light has been used and studied in the past for numerous applications, the understanding of the polarization phenomenology taking place with targets used in cluttered backgrounds requires additional experimentations. Specifically, the target contrast enhancement obtained by analyzing the polarization of the reflected light from either a direct polarized laser source as encountered in active imagers, or from natural ambient illumination, needs further investigation. This paper describes an investigation of the use of polarization-based imaging sensors to discriminate civilian and military targets against different backgrounds. Measurements were carried out using two custom-designed active and passive imaging systems operating in the near infrared (NIR) and the long-wave infrared (LWIR) spectral bands. Polarimetric signatures were acquired during two distinct trials that occurred in 2007, using specific civilian and military targets such as cars and military vehicles. Results demonstrate to what extent and under which illumination and environmental conditions the exploitation of active and passive polarimetric images is suitable to enable target detection and recognition for some events of interest, according to various specific scenarios.

Robust component-based car detection in aerial imagery with new segmentation techniques

Several new techniques are introduced to the component-based vehicle detection in the aerial imagery. The shape-independent tricolour attenuation model based on the spectral power density difference between the regions lighted by direct sunlight and/or diffuse skylight is used to identify cast shadows. The simple linear iterative clustering (SLIC) performs local clustering for superpixels, which were merged by a statistical region merging (SRM) method based on the independent bounded difference inequality theorem. The car body parts were found with Support Vector Machine based on the radiometric and geometric features of the segmented regions. All the algorithms used in this approach require minimum human intervention, providing a robust detection.

Development of performance metrics to characterize the degree of polarization of man-made objects using passive polarimetric images

Spectral sensors are commonly used to measure the intensity of optical radiation and to provide spectral information about the distribution of material components in a given scene, over a limited number of wave bands. By exploiting the polarization of light to measure information about the vector nature of the optical field across a scene, collected polarimetric images have the potential to provide additional information about the shape, shading, roughness, and surface features of targets of interest. The overall performance of target detection algorithms could thus be increased by exploiting these polarimetric signatures to discriminate man-made objects against different natural backgrounds. This is achieved through the use of performance metrics, derived from the computed Stokes parameters, defining the degree of polarization of man-made objects. This paper describes performance metrics that have been developed to optimize the image acquisition of selected polarization angle and degree of linear polarization, by using the Poincare sphere and Stokes vectors from previously acquired images, and then by extracting some specific features from the polarimetric images. Polarimetric signatures of man-made objects have been acquired using a passive polarimetric imaging sensor developed at DRDC Valcartier. The sensor operates concomitantly (bore-sighted images, aligned polarizations) in the visible, shortwave infrared, midwave infrared, and the long-wave infrared bands. Results demonstrate the improvement of using these performance metrics to characterize the degree of polarization of man-made objects using passive polarimetric images.

Robust vehicle detection in aerial images based on salient region selection and superpixel classification

For detecting vehicles in large scale aerial images we first used a non-parametric method proposed recently by Rosin to define the regions of interest, where the vehicles appear with dense edges. The saliency map is a sum of distance transforms (DT) of a set of edges maps, which are obtained by a threshold decomposition of the gradient image with a set of thresholds. A binary mask for highlighting the regions of interest is then obtained by a moment-preserving thresholding of the normalized saliency map. Secondly, the regions of interest were over-segmented by the SLIC superpixels proposed recently by Achanta et al. to cluster pixels into the color constancy sub-regions. In the aerial images of 11.2 cm/pixel resolution, the vehicles in general do not exceed 20 x 40 pixels. We introduced a size constraint to guarantee no superpixels exceed the size of a vehicle. The superpixels were then classified to vehicle or non-vehicle by the Support Vector Machine (SVM), in which the Scale Invariant Feature Transform (SIFT) features and the Linear Binary Pattern (LBP) texture features were used. Both features were extracted at two scales with two size patches. The small patches capture local structures and the larger patches include the neighborhood information. Preliminary results show a significant gain in the detection. The vehicles were detected with a dense concentration of the vehicle-class superpixels. Even dark color cars were successfully detected. A validation process will follow to reduce the presence of isolated false alarms in the background.

Comparing several AFE tools in the context of ships and vehicles detection based on RGB and EO data

In this paper we present a comparison of three AFE tools used in the context of ship and vehicle detection based on high resolution data. The three tools (Genie Pro - Los Alamos National Laboratory, Feature Analyst - Visual Learning Systems and eCognition - Definiens AG) were chosen because they were defined as promising and were to be analyzed by NGA in the framework of the STAR program. The comparison is presented here in terms of detection and false alarm rates and also in terms of pros and cons of each tool.

A fully automated image co-registration system

Current surveillance and reconnaissance systems require improved capability to enable the co-registration of larger images, combining enhanced temporal, spatial, and spectral resolutions. However, such proficient remote sensing systems cannot employ traditional manual exploitation techniques to cope successfully with the avalanche of data to be processed and analyzed. Automated image exploitation tools may be employed if the images are already co-registered together. Therefore, there is a need to develop fully automated co-registration algorithms able to deal with different scenarios, and helpful to be used successively for numerous applications such as image data fusion, change detection, and target detection. This paper describes the Automated Multi-sensor Image Registration (AMIR) system and embedded algorithms under development at DRDC-Valcartier. The AMIR system provides a framework for the automated multi-date registration of electro-optic images, acquired from different sensors and from dissimilar oblique view angles. The system is characterized by its fully automated nature, where no user intervention prevailed. Advanced image algorithms are used in order to supply the capability to register multi-date electro-optic images acquired from different viewpoints, under singular operational conditions, multiple scenarios (e.g. airport, harbor, vegetation, urban, etc.), different spatial resolutions (e.g. IKONOS/QuickBird, Airborne/Spaceborne), while providing sub-pixel accuracy registration level.