Robust detection of preceding vehicles in crowded traffic conditions

This paper introduces an adaptive scheme for robustly detecting multiple preceding vehicles in crowded traffic conditions. The scheme focuses on issues frequently observed in the interpretation of traffic scenes recorded by cameras installed in vehicles: stable extraction of features and accurate classification in spite of the vehicle??s constant vibrations and dynamic changes in the distance between vehicles. To address these issues, we introduce the concept of integral features and a method of utilizing the scene geometry information. Each of the simple attributes, such as edges, shadows, and symmetry, is compiled in the window confined by the scene geometry to improve the expressiveness and robustness of the extracted features. The scene geometry information that is then estimated from the perspective view is extensively utilized in constituent system components, including not only feature extraction/integration but also neural network-based classification and distance-adaptive clustering. In addition, employing the Kalman filter along with a confidence measure makes the detection and tracking of potential vehicles robust. Experimental results prove that the system employing the proposed scheme detects and tracks multiple vehicles more effectively, even in crowded traffic conditions, with a lower rate of false positives.