Research And Implementation Of A Dynamic Feature Selection Method For Vehicle Recognition System

Effective and accurate detection and tracking of moving vehicles in video sequences is the key to modern intelligent traffic monitoring systems. Vision based vehicle recognition has received extensive attentions because of its fine applicability and high cost-performance ratio. Thus, vision based vehicle recognition has become a popular research topic in the field of image processing and artificial intelligence. The recognition algorithm mainly implements the synthesis which uses both appearance based and knowledge based features to identify its candidates. However, due to the unpredicted complex noises in real world environments, existences, quantifications and explanations for certain features are often ambiguous, which makes current algorithm hard to fulfill the dilemmatic high sensitivity/accuracy restriction, and an improvement for a certain feature (or data sets) often leads to degeneration for others. This dissertation provides a mathematical model for vehicle recognition, and proposes a probability model based feature selection method which enables the dynamic feature selection and multigrain feature evaluations. The method of feature digraph construction takes very sufficient thought of the fact that, in the real world, the imbalance that different features result in different outcomes of recognitions; and the inter-dependencies and the inter-relationships among different features, etc. By doing so it makes the entire recognition process converge to the objective laws that vehicle recognitions is of multi-feature matching and collaboration. It is the more accurate and more effective imitation of the process of human-recognitions. Eventually, by comparing with the former rear vehicle recognition algorithms in experiments, ROC curves show that not only the recognition rate was improved in a variety of thresholds, but also the accuracy was raised from 0.9286 to 0.9513, it further proves that the method (mentioned above) is an effective solution which can increase the accuracy and sensitivity of recognitions, and it is very helpful in solving feature-racing alike problems.