Method of Multi-lane Vehicles Speed Continuously Perceiving Based on Single Roadside Camera

Roadside camera has been widely applied to detect the traffic status and now it is an important component composing the digital road infrastructure. A novel method of multi-lane vehicles speed continuously perceiving based on single roadside camera is proposed in this paper. Firstly, extended Haar feature is adopted by identifying objects of roadside camera video to achieve the training data set. Then, an AdaBoost cascade classifier is designed and optimized based on iterative learning of the data set for accurately vehicle identifying. Thirdly, an association tracker is proposed based on MOSSE to realize multi-vehicle tracking in consecutive video frames, and average pixel and Euclidean distance are applied to locate the vehicle position and calculate the vehicle trajectory. At last, a transformation relation of image pixel to physical distance is proposed to obtain the vehicle real time speed. The proposed method has been verified with real roadside camera data. The experimental results show that the vehicle recognizing accuracy is above 98.02%, the vehicle speed perceiving error is within $\pm 2\%$, and the proposed method can deal with real time roadside camera data with good.