Vehicle Dimensions and Speed Estimation using Camera Calibration Based on Recognition of a Number of Common Cars by VGG Network

This paper presents an automated method for calibrating road cameras in order to estimate the speed and dimensions of the vehicles. In this method, in the initial frames, according to the direction of vehicles movement, vanishing points and the hypothetical road surface is obtained. Then, by identifying the foreground using IGMM and removing the shadow, the exact boundary of each vehicle is determined and a 3D bounding box is constructed. To determine the metric coefficients, several vehicles from common classes are identified using the deep VGG neural network in the first few frames. Further, according to the actual dimensions of the vehicles identified in meters and their equivalent dimensions on the road surface in pixels, the metric coefficients are calculated and the calibration parameters are completed. Ultimately, passing cars are projected on the hypothetical page, and by tracking each car, its speed and dimension are calculated. A database of vehicle images was collected to identify common cars. To evaluate our method, a series of videos with ground truth was provided, by simultaneous capture of road vehicles by RGB and laser camera. The mean error of the proposed method for speed estimating is 1.15 km /h and for dimension estimation is equal to 2.3%, which shows good performance of the method.