Automatic adaptation in classification algorithms fusing data from heterogeneous sensors

In a heterogeneous multisensor environment, data fusion can help to improve the reliability of the sensor data. With respect to classification of tracked objects, features detected by different types of sensors can supplement each other, creating more reliable classifications results. This paper presents an idea of adapting the classification rules in a learning process without "known good" learning data. Instead, the fused result of classification by different classifiers is used for learning. This way, the classification rules are refined in absence of a prior knowledge. The gain lies in the fact that the fused classification results from heterogeneous sensors can be considered a good approximation of the absolute truth. In an urban scenario there exist many areas with different sensor coverage. The learning is most effective in areas viewed by many sensors - the fused result is more reliable there. Rules adapted using objects present in these areas can then be successfully used in the areas where only one sensor is active. A DAFNE fusion engine together with DAFNE sensor simulator is described in the paper as a platform which enables experimenting with different fusion algorithms. Results of experiments with the proposed algorithm and Naïve Bayes Classifier are shown. The experiments were designed to show the learning process itself and to study the quality of learning in the presence of disturbed sensor readings.