Evaluation of Tracking Confidence Indicators and Feature Extractors on a Visual Tracking Algorithm

Abstract

For visual tracking, a radial basis function neural network algorithm will be used. Coupled with a feature extraction algorithm, the neural network has advantages for pattern recognition, including practical implementation in parallel hardware for real-time operation and low power requirements. Targets vary in terms of texture, contrast, sharpness of edge, relative speed, and size. Various feature extractors exhibit tradeoffs in terms of sensitivity and processing requirements as related to the characteristics of candidate target classes. An analysis of feature extractors based on the horizontal and vertical profile has been provided. A comparison of the distance traveled computed from vision to wheel encoders is presented to observe slipping. Feedback from the network can offer an indication of tracking confidence which will be useful in determining if the estimated position is correct. An attempt has been made to look at the various confidence factors to determine if the position estimated is correct.