Improving the performance of distant object classification with advanced pre-processing

Object classification is an important topic in many surveillance and reconnaissance applications. False detections can be suppressed, potentially suspicious objects identified, and finally different object types separated. However, robust classification is difficult to accomplish as objects don't operate cooperatively, object distance may be high, and various sensor specific noise-effects are to be handled. Appropriate pre-processing can be very helpful for the classification process. This ranges from standard noise filters to advanced methods achieving scale- and rotation-invariance, which significantly supports the classifier performance and generality. In this work, several advanced and noise-resistant methods are presented with respect to three pre-processing tasks: scale-invariance, rotation-invariance, and precise object segmentation. The benefit of these methods is demonstrated using several bird's eye view real data examples coming from different imaging sensors.