Определение характеристик движущихся объектов с использованием "контролируемого" размытия изображения

Abstract

The paper proposes an approach to detecting, capturing and identifying vehicles in an image in order to determine their geometrical, kinematic and dynamic properties. Monitoring the properties determined in real time will assist in detecting violations of vehicle regulations, in analysing and preventing potential accidents by means of using the parameters obtained to control vehicle motion, as well as in implementing a digital model of the environment, including its possible virtual reality representation. To detect the object, we employ the YOLOv3 algorithm using convolutional neural network training. Canny and Hough detectors are used to determine the object boundaries and its axle positions. We propose a method for determining the object boundary blurring based on numerically finding the first, second and third derivatives with respect to each coordinate axis and subsequently refining the result at different resolutions. After detecting the axles in a vehicle and determining their number via passive image analysis, a model of vertical oscillations is selected for the vehicle; the geometrical and kinematic parameters obtained are then used to assess the automotive vehicle dynamics. We describe in detail brour vehicle model featuring two axles. We assume the parameters of an artificial obstacle such as a speed breaker to be the initial conditions of the oscillating system