Inductive loop sensors are commonly used to detect traffic violations. However

these tools are expensive and difficult to maintain. Tracing vehicles or detecting violation by video analysis represents an alternative way to exploit computer vision. Thus

the calibration of intersection driveways

which is the basic task of the intelligent transportation system in the detecting traffic violations

must be primarily addressed. This paper presents a solution for extracting intersection backgrounds and marking out driveways. First

we propose a new background extraction algorithm

which inherits the features of both mean and frame difference methods

to solve the problem of lane markers that are partly covered by vehicles. On the one hand

this algorithm exploits the average image to estimate an extra multiplying power that keeps the background relatively stable. On the other hand

this algorithm calculates the frame difference and uses such difference to update the background progressively. Thus

the proposed algorithm achieves fast convergence when the traffic runs smoothly and can be quickly updated during traffic hours. Second

several lines are detected by employing Canny and Hough. Based on perspective transformation

clustering analysis

and prior knowledge

a filtering mathematical model is proposed to detect the driveway from these lines. Third

the proposed algorithm is verified by conducting experiments. The proposed algorithm can obtain a more robust outcome than the Gaussian mixture model with five Gaussian distributions

which is one of the most widely used background extraction methods. Using the manual background as ground truth

the proposed algorithm can quantitatively compare the gray value of the ground truth's pixel with the corresponding extracted background's pixel. The result of the experiments shows that the accuracy rate of background extraction is 20% and 30% more than that of the mean method and the traditional Gaussian mixed model

respectively. According to the cycle data of traffic lights

the ghost can be avoided when the vehicles stop at red light. In other words

the proposed algorithm can distinguish between a temporary stopped objective and a long-time stationary background. Similarly

the calibration method can precisely determine the pseudo lane lines using a clustering or filtering strategy and produce a reliable result. The proposed algorithm has several merits

such as fast convergence rate

higher accuracy rate

and excellent stability

which can rapidly erase the virtual shadow of the vehicle. In addition

calibration can be accomplished in the daytime as a one-time effort. Thus

the proposed algorithm is suitable to calibrate driveways with normal lights. The experiments also demonstrate the effectiveness and practicality of the proposed method. However

the algorithm still requires further optimization and analysis in adjusting parameters. In the future

we will attempt to devise a new method to select the suitable parameters adaptively by adopting some machine learning approaches.