Abstract: Described is system and method for visual media reasoning. An input image is filtered using a first series of kernels tuned to represent objects of general categories, followed by a second series of sparse coding filter kernels tuned to represent objects of specialized categories, resulting in a set of sparse codes. Object recognition is performed on the set of sparse codes to generate object and semantic labels for the set of sparse codes. Pattern completion is performed on the object and semantic labels to recall relevant meta-data in the input image. Bi-directional feedback is used to fuse the input data with the relevant meta-data. An annotated image with information related to who is in the input image, what is in the input image, when the input image was captured, and where the input image was captured is generated.

Abstract: Described is a system for detecting moving objects using multi-frame motion history images. An input video sequence of consecutive registered image frames is received. The sequence of consecutive registered image frames comprises forward and backward registered image frames registered to a coordinate system of a reference image frame. Frame differences are computed between each of the consecutive registered image frames and the reference image frame. The frame differences are accumulated based on characteristics of the input video sequence to compute a corresponding motion response value. A selected threshold value is then applied to the motion response value to produce at least one binary image used for detection of moving objects in the input video sequence. Additionally, the invention includes a system for adaptive parameter optimization by input image characterization, wherein parameters that are based on characteristics of the image influence the motion detection process.

Abstract: The present invention relates to a classifier cascade object detection system. The system operates by inputting an image patch into parallel feature generation modules, each of the feature generation modules operable for extracting features from the image patch. The features are provided to an opportunistic classifier cascade, the opportunistic classifier cascade having a series of classifier stages. The opportunistic classifier cascade is executed by progressively evaluating, in each classifier in the classifier cascade, the features to produce a response, with each response progressively utilized by a decision function to generate a stage response for each classifier stage. If each stage response exceeds a stage threshold then the image patch is classified as a target object, and if the stage response from any of the decision functions does not exceed the stage threshold, then the image patch is classified as a non-target object.

Abstract: The present invention relates to a system for detecting an object of interest in a scene. The system operates by receiving an image frame of a scene and extracting features from the image frame, the features being descriptors. The descriptors are quantized to generate PHOW features. A sliding window protocol is implemented to slide a window over the image and analyze the PHOW features that fall inside the window. Finally, the system determines if the PHOW features represent the object of interest and, if so, then designates the window as a location in the image with a detected object of interest.

Abstract: Described is a system for improving object recognition. Object detection results and classification results for a sequence of image frames are received as input. Each object detection result is represented by a detection box and each classification result is represented by an object label corresponding to the object detection result. A pseudo-tracklet is formed by linking object detection results representing the same object in consecutive image frames. The system determines whether there are any inconsistent object labels or missing object detection results in the pseudo-tracklet. Finally, the object detection results and the classification results are improved by correcting any inconsistent object labels and missing object detection results.

Abstract: Described, is a system for object detection via multi-scale attentional mechanisms. The system receives a multi-band image as input. Anti-aliasing and downsampling processes are performed to reduce the size of the multi-band image. Targeted contrast enhancement is performed on the multi-band image to enhance a target color of interest. A response map for each target color of interest is generated, and each response map is independently processed to generate a saliency map. The saliency map is converted into a set of detections representing potential objects of interest, wherein each detection is associated with parameters, such as position parameters, size parameters, an orientation parameter, and a score parameter. A post-processing step is applied to filter out false alarm detections in the set of detections, resulting in a final set of detections. Finally, the final set of detections and their associated parameters representing objects of interest is output.

Abstract: Described is system and method for robust ground-plane homography estimation using adaptive feature selection. The system determines feature correspondences of an image that correspond with at least one moving object in each image in a set of images. Additionally, feature correspondences of the image that correspond with at least one above-ground object are determined in each image. Feature correspondences that correspond with each moving object in each image are excluded, and feature correspondences that correspond with each above-ground object in each image are excluded. Each image is divided into a plurality of sub-regions comprising features correspondences. The number of feature correspondences in each sub-region is limited to a predetermined threshold to ensure that feature correspondences are evenly distributed over each image. Finally, a ground-plane homography estimation between the set of images is generated.

Abstract: Described is a system for multi-object detection and recognition in cluttered scenes. The system receives an image patch containing multiple objects of interest as input. The system evaluates a likelihood of existence of an object of interest in each sub-window of a set of overlapping sub-windows. A confidence map having confidence values corresponding to the sub-windows is generated. A non-maxima suppression technique is applied to the confidence map to eliminate sub-windows having confidence values below a local maximum confidence value. A global maximum confidence value is determined for a sub-window corresponding to a location of an instance of an object of interest in the image patch. The sub-window corresponding to the location of the instance of the object of interest is removed from the confidence map. The system iterates until a predetermined stopping criteria is met. Finally, detection information related to multiple instances of the object of interest is output.

Abstract: Described is a system for rapid object detection combining structural information with bio-inspired attentional mechanisms. The system oversegments an input image into a set of superpixels, where each superpixel comprises a plurality of pixels. For each superpixel, a bounding box defining a region of the input image representing a detection hypothesis is determined. An average residual saliency (ARS) is calculated for the plurality of pixels belonging to each superpixel. Each detection hypothesis that is out of a range of a predetermined threshold value for object size is eliminated. Next, each remaining detection hypothesis having an ARS below a predetermined threshold value is eliminated. Then, color contrast is calculated for the region defined by the bounding box for each remaining detection hypothesis. Each detection hypothesis having a color contrast below a predetermined threshold is eliminated. Finally, the remaining detection hypotheses are output to a classifier for object recognition.

Abstract: Described is a method and system for embedding unsupervised learning into three critical processing stages of the spatio-temporal visual stream. The system first receives input video comprising input video pixels representing at least one action and at least one object having a location. Microactions are generated from the input image using a set of motion sensitive filters. A relationship between the input video pixels and the microactions is then learned, and a set of spatio-temporal concepts is learned from the microactions. The system then learns to acquire new knowledge from the spatio-temporal concepts using mental imagery processes. Finally, a visual output is presented to a user based on the learned set of spatio-temporal concepts and the new knowledge to aid the user in visually comprehending the at least one action in the input video.

Abstract: Described is a system for selective color processing for vision systems. The system receives a multi-band image as input. As an optional step, the multi-band image is preprocessed, and a transformation is performed to transform the multi-band image into a color space. A metric function is applied to the transformed image to generate a distance map comprising intensities which vary based on a similarity between an intensity of a pixel color and an intensity of a target color. A contrast enhancement process is applied to the distance map to normalize the distance map to a range of values. The range of values is expanded near the intensity of the target color. Finally, an output response map for the target color of interest is generated, such that the output response map has high responses in regions which are similar to the target color to aid in detection and recognition processes.

Abstract: Described is method for object cueing in motion imagery. Key points and features are extracted from motion imagery, and features between consecutive image frames of the motion imagery are compared to identify similar image frames. A candidate set of matching keypoints is generated by matching keypoints between the similar image frames. A ground plane homography model that fits the candidate set of matching keypoints is determined to generate a set of correct matching keypoints. Each image frame of a set of image frames within a selected time window is registered into a reference frame's coordinate system using the homography transformation. A difference image is obtained between the reference frame and each registered image frame, resulting in multiple difference images. The difference images are then accumulated to calculate a detection image which is used for detection of salient regions. Object cues for surveillance use are produced based on the detected salient regions.

Abstract: Described is a system for object detection using classification-based learning. A fusion method is selected, then a video sequence is processed to generate detections for each frame, wherein a detection is a representation of an object candidate. The detections are fused to generate a set of fused detections for each frame. The classification module generates a classification score labeling each fused detection based on a predetermined classification threshold. Otherwise, a token indicating that the classification module has abstained from generating a classification score is generated. The scoring module produces a confidence score for each fused detection based on a set of learned parameters from the learning module and the set of fused detections. The set of fused detections are filtered by the accept-reject module based on one of the classification score or the confidence score. Finally, a set of final detections representing an object is output.

Abstract: A method and apparatus for processing images. A sequence of images for a scene is received from an imaging system. An object in the scene is detected using the sequence of images. A viewpoint of the imaging system is registered to a model of the scene using a region in the model of the scene in which an expected behavior of the object is expected to occur.

Abstract: Described is a system for stabilizing, detecting, and recognizing objects in video captured from a mobile platform. The system first receives a video (with a plurality of image frames) captured from a mobile platform. The video is stabilized by registering the image frames to a global coordinate system to generate stabilized image frames. A bio-inspired attention algorithm is applied to the stabilized image frames to produce a set of locations in the stabilized image frames that are salient points representative of an object of interest. An image chip is generated that surrounds each salient point. High-dimensional feature vectors are extracted from the image chip. The feature vectors are then classified as an object class. Thus, through classifying the feature vectors, an object of interest can be identified in the video as captured from the mobile platform.

Abstract: Described is a system for registering a viewpoint of an imaging sensor with respect to a geospatial model or map. An image of a scene of a geospatial region comprising an object is received as input. The image of the scene is captured by a sensor having a current sensor state. Observation data related to the object's state is received, wherein the observation data comprises an object behavior of the object given the geospatial region. An estimate of the current sensor state is generated using a probability of an observation from the observation data given the current sensor state x. Finally, the image of the scene is registered with a geospatial model or map based on the estimate of the current sensor state.

Abstract: Described is a system for multi-class classifier threshold-offset estimation for visual object recognition. The system receives an input image with input features for classifying. A pair-wise classifier is trained for each pair of a plurality of object classes. A set of classification responses is generated, and a multi-class receiver-operating-characteristics (ROC) curve is computed for a set of threshold-offsets. An objective function of classification performance is computed from the ROC curve and optimized using particle swarm optimization (PSO) to generate a set of optimized threshold-offsets. The optimized threshold-offsets are then applied to the classification responses. The resulting classification responses are compared to a predetermined value to classify each input feature as belonging to one object class or another. The tuning of the threshold-offsets with (PSO) improves classification performance in a visual object recognition system.

Abstract: A method and apparatus for processing images. A sequence of images for a scene is received from an imaging system. An object in the scene is detected using the sequence of images. A viewpoint of the imaging system is registered to a model of the scene using a region in the model of the scene in which an expected behavior of the object is expected to occur.