Robust real-time change detection in high jitter.

Abstract

A new method is introduced for real-time detection of transient change in scenes observed by staring sensors that are subject to platform jitter, pixel defects, variable focus, and other real-world challenges. The approach uses flexible statistical models for the scene background and its variability, which are continually updated to track gradual drift in the sensor's performance and the scene under observation. Two separate models represent temporal and spatial variations in pixel intensity. For the temporal model, each new frame is projected into a low-dimensional subspace designed to capture the behavior of the frame data over a recent observation window. Per-pixel temporal standard deviation estimates are based on projection residuals. The second approach employs a simple representation of jitter to generate pixelwise moment estimates from a single frame. These estimates rely on spatial characteristics of the scene, and are used gauge each pixel's susceptibility to jitter. The temporal model handles pixels that are naturally variable due to sensor noise or moving scene elements, along with jitter displacements comparable to those observed in the recent past. The spatial model captures jitter-induced changes that may not have been seen previously. Change is declared in pixels whose current values are inconsistent with both models.

Cell Multi-Bernoulli (Cell-MB) Sensor Control for Multi-Object Search-While-Tracking (SWT)

“…Object detections are generated from raw frame data using normalized difference change detection [32] and fast approximate power iteration subspace tracking [33] for temporal background estimation. The single-object measurement function is linear-Gaussian with corresponding likelihood g(z|x) = N (z; Hx, R) , (77)…”

Cell Multi-Bernoulli (Cell-MB) Sensor Control for Multi-object Search-While-Tracking (SWT)

“…Object detections are generated from raw frame data using normalized difference change detection [30] and fast approximate power iteration subspace tracking [31] for tempo-ral background estimation. The single-object measurement function is linear-Gaussian with corresponding likelihood…”

A Variational Approach for Sparse Component Estimation and Low-Rank Matrix Recovery

“…However, Algorithm 11.2, as well as most RPCA algorithms, have a high sensitivity to camera jitter, which can affect airborne and space-based sensors [53] as well as fixed groundbased cameras [10] subject to wind. Algorithm 11.2 does the job, eliminating the batch processing mode, typically needed for RPCA algorithms whereby a large number of frames have to be observed before starting any processing.…”