Probabilistic Model and Neural Network for Scene Classification in Traffic Surveillance System

Traffic surveillance system (TSS) has seen great progress in the last several years. Many algorithms have been developed to cope with a wide range of scenarios such as overcast, sunny weather that created shadows, rainy days that result in mirror reflection on the road, or nighttime when low lighting conditions limit the visual range. However, in real-world applications, one of the most challenging problems is the scene determination in a highly dynamic outdoor environment. As also pointed out in recent survey, there have been limited studies on a mechanism for scene recognition and adapting appropriate algorithms for that scene. Therefore, this research presents a scene recognition algorithm for all-day surveillance. The proposed method detects and classifies outdoor surveillance scenes into four common types: overcast, clear sky, rain, and nighttime. The major contributions are to help diminish hand-operated adjustment and increase the speed of responding to the change of alfresco environment in the practical system. To obtain high reliable results, we combine the histogram features on RGB color space with the probabilistic model on CIE-Lab color space and input them into a feedforward neural network. Early experiments have suggested promising results on real-world video data.