Integrating Temporal Streams of Image Classifications Using Evidential Reasoning for Smart Airbag Applications

In many real-time classification applications, a sequence of images are collected and classified over time. During operation, there may be times when the system may experience a variety of situations that may make the correct classification difficult or even impossible. These changes result in a new type of classification errors, namely assignable errors, rather than random errors. We propose a framework based on the theory of evidential reasoning to process a real-time sequence of image classifications results. We develop a general framework based on evidential reasoning and then demonstrate its effectiveness using Bayesian and Dempster-Shafer inference. We then compare their performance with traditional classifier combination algorithms on the application of a real-time vision system for smart automotive airbags. We show that the evidential reasoning approaches outperform traditional classifier combination method by roughly 10% and provide roughly a 20% improvement over single image decisions