SEGMENTATION FREE SHARED WEIGHT NETWORKS FOR AUTOMATIC VEHICLE DETECTION

The problem of detecting a vehicle of a certain type in an outdoor sce ne is investigated. A shared weight neural network is applied in corre lation fashion to images of outdoor scenes. Two sets of images are con sidered. The first set consists of forward looking infrared images of tanks. The second set consists of visible images of cars in a parking lot. The goal in the first set of images is to detect the tanks. The g oal in the second set of images is to detect the Chevrolet Blazers. Th e second problem is much more difficult than the first. The networks a re compared to a minimum average correlation energy (MACE) matched fil ter approach. The shared weight network was found to be better at gene ralizing. Problems were found in trying to suppress spurious outputs o n the background. As a result, an improved training algorithm was deve loped.