Signal processing and object recognition in 3-D laser imagery on a Connection Machine

Summary form only given. Imagery from an aircraft laser sensor is being processed to find camouflaged and uncamouflaged vehicles in varying backgrounds. The laser sensor operates at 0.85 μm and measures, using cross-track scanning, both range from the sensor and reflectivity of the scene at a resolution of 6 in. For a typical application, the range and reflectance imagery is produced at a rate of 4 million pixels/s. The processing, done on a Connection Machine, a fine-grained, massively parallel supercomputer, includes two steps: removing artifacts in the raw signal data and automatically recognizing objects using a hybrid statistical classification and model-matching approach. The artifact removed is a sawtooth pattern superimposed on the range information caused by the 10-m range ambiguity of the sensor. Object recognition occurs as a two-step process: statistical classification for rapid image screening and object detection, and 3-D model matching for object recognition. A highly parallel model matching scheme is used to complete the recognition process