Abstract
Novel algorithms for detecting, classifying and tracking vehicles in infra red images are presented and their implementation on an array of transputers is discussed. Two classification techniques have been developed; the first, Segmentation Analysis, partitions each image into rectangular subimages from which several feature parameters are extracted. It is hypothesised that those subimages which have features differing significantly from the mean feature values will contain objects of interest which can be classified from the values of the features. In the second technique, ACF Classification, simple parameters are derived from the shape of the autocorrelation functions from a segmented image. The parameters are then be used to assign the segmented regions to the appropriate class of target or background. The ACF classification algorithm has been tested against a standard data base and has given a performance comparable with a human observer. A computationally efficient tracking algorithm has been designed which matches histograms from successive frames with a histogram from a reference image giving significant advantages in execution time over conventional algorithms. Further improvement in execution time can be obtained by combining histogram matching with novel search patterns. The algorithms were initially implemented on a microprocessor based image processor to demonstrate feasibility. Following a detailed emulation of the Inmos transputer, the algorithms were implemented on a multi-transputer network to demonstrate real-time image processing.