Efficient intelligent transportation systems (ITS) should be built on the real-time traffic information collection. With the deep research of intelligent video surveillance, video-based traffic information collection attracts attentions of numerous researchers since its broad application prospects. Then, the technique to detect traffic objects in images is the core and foundation for information collection and other high-level tasks. Currently, with the development of image acquisition devices, traditional analog cameras with D1 resolutions are gradually replaced with high-definition digital video cameras, which have been mainstream devices in traffic application. New opportunities may be brought to solve many fixed difficulties in object detection because of the high-definition characteristics. Based on complex video surveillance scenarios in urban traffic environments, we select two representative objects, license plates and pedestrians, to achieve their detection with the help of computer vision, machine learning, and other technologies. We use component-based detection methods and introduce probabilistic graphical models (PGM). Efforts are made to solve many common difficulties in object detection, such as occlusion, large intra-class variations, cluttered background. The main work and contributions of this paper are: (1) A license plate detection algorithm based on nearest neighbor chain (NNC) of connected components is proposed to detect license plates with characters aligning in chain structures. The component-based object detection is utilized in this task to use both appearance features and structural features of license plates. Firstly, the maximally stable extremal region (MSER) detector is introduced to extract stable connected components as candidate license plate characters in the image. Then, the nearest neighbor pairs (NNP) of connected components are constructed with the spatial relationships to build NNCs. Finally, sequence encoding and matching strategies are performed on NNCs. Experimental results show the effectiveness of the algorithm in various real traffic environments with extremely low false positive rate. (2) A license plate detection algorithm using conditional random field (CRF) models is proposed. In order to solve the limitations of NNC-based license plate algorithms, PGM are introduced to make the algorithm adapt to more complex character arrangement and more challenging environments. MSERs that meet certain neighborhood...
修改评论