中央研究院 資訊科學研究所

In this work, we propose a self-diagnosing intelligent highway surveillance system. The system has the ability to perform self-diagnosis and choose the most suitable algorithm for current conditions. Daytime and nighttime modules are developed. In particular, a more efficient and effective headlight pairing algorithm is designed for the nighttime module. By analyzing the camera geometry, the distances between headlight pairs at different positions could be obtained without prior assumptions of many camera parameters. For surveillance videos under different camera angles, a major advantage of this mechanism is that there is no need to recollect training vehicle samples. For different surveillance scenes, instead of recollecting the vehicle image samples and retrain the classifiers, the proposed method only requires re-measuring the pixel distance between headlights at a certain position of the scene, which would be much easier. Furthermore, this system balances between abundance of acquired information and robustness. Performing tracking would be preferred when traffic is smooth in order to acquire useful information for subsequent event analysis modules. However, under severe congestion conditions, estimation via regression is performed to obtain traffic parameters because tracking results would be unreliable under such circumstances.

 

We also discuss the concept of self-diagnosis to detect the conditions when the camera is seriously tampered by rain-drops. We provide solutions to analyze the traffic flow under the challenging rain-drop tampered conditions. To deal with the challenging scenes, effective features are extracted via salient region detection and block segmentation. The extracted features are used to train a support vector machine to achieve self-diagnosis. For traffic flow analysis, we use the extracted features in the region of interest and construct a regression model to get an estimated vehicle count for each frame. The vehicle counts in consecutive frames form a vehicle count sequence. We propose a mapping model to acquire the desired per minute traffic flow from the vehicle count sequence. The model utilizes state transfer likelihoods and takes into account the length of the segmented vehicle count sequence. With highly challenging data sets, we have demonstrated that the proposed system can balance between abundance of acquired information and robustness. For conditions under which tracking is not feasible, the system can effectively estimate the traffic flow.