Abstract: Described is a low power surveillance camera system for intruder detection. The system observes a scene with a known camera motion to generate images with various viewing angles. Next, a background learning mode is employed to generate a low rank matrix for the background in the images. Background null space projections are then learned, which provide a foreground detection kernel. A new scene with known viewing angles is then obtained. Based on the foreground detection kernel and the new input image frame, low power foreground detection is performed to detect foreground potential regions of interest (ROIs), such as intruders. To filter out minimal foreground activity, the system identifies contiguous ROIs to generate the foreground ROI. Focus measures are then employed on the ROIs using foveated compressed sensing to generate foveated measurements. Based on the foveated measurements, the foreground is reconstructed for presentation to a user.

Abstract: A back end-circuit for randomized non uniform and alias-free subsampling, comprising: an analog-to-digital converter (ADC) configured for sampling an input signal at random non uniform times; a compressive sensing processor, coupled to the ADC, to recover a sparse spectral representation of the input signal; and a Fourier transformer for converting the sparse spectral representation to a time sampled representation of the input signal.

Abstract: Described is a system for foveated compressive sensing. The system is configured to receive an input image f of a scene and initialize a measurement matrix. Global measurements are then performed, with a lower resolution image of the scene thereafter reconstructed. Task salient regions are extracted from the low resolution image. Thereafter, the system estimates a task-specific operator and detects regions-of-interest (ROI) based on the task salient regions. An ROI-adapted and foveated measurement matrix is then generated. Local measurements are then performed on task-relevant ROIs. A higher resolution image can then be reconstructed of the scene to allow for identification of objects in the ROI.

Abstract: The present invention relates to an object detection and behavior recognition system using three-dimensional motion data. The system receives three-dimensional (3D) motion data of a scene from at least one sensor, such as a LIDAR sensor. An object is identified in the 3D motion data. Thereafter, an object track is extracted, the object track being indicative of object motion in the scene over time. Through Dynamic Time Warping (DTW) or other comparison techniques, the object track is compared to a database to identify the behavior of the object based on its object track.