An heterogeneous M-SIMD architecture for Kalman Filter controlled processing of image sequences

An heterogeneous Multiple-SIMD (M-SIMD) architecture is used to analyse image sequences by integrating image processing operations with optimal recursive (Kalman) estimators. The architecture uses SIMD processors for data parallel (iconic) operations and MIMD processors for control parallel (numeric and symbolic) tasks. The SIMD processors are configured as small contiguous sub-arrays, with each subarray attached to a single MIMD processor. This allows operational autonomy intermediate between the pure data and control parallel paradigms. Use of the architecture is illustrated with size-based detection and segmentation techniques that are guided by Kalman filters through an image sequence. The measurements of object size on the image plane allow a robust estimate of the range and physical size of the object in the world. Results are presented for the architecture performance, the image processing, and the estimation techniques