Vehicle Speed and Dimensions Estimation from On-Road Cameras by Identifying Popular Vehicles

Abstract

One of the major issues related to traffic monitoring systems is analyzing the behavior of vehicles and identifying their characteristics. In this paper, an automated algorithm is proposed for calibration of the road cameras. This calibration is used to estimate the speed and dimensions of the passing vehicles. In this method, a motion plane is obtained in the initial frames according to the direction of moving vehicles. After modeling the background, removing shadows, and identifying the exact boundaries of passing vehicles, a number of common vehicles are recognized by the Bag of Words method. Given the actual metric dimensions of these vehicles and the equivalent dimensions on the motion plane, the metric coefficients are calculated and the calibration process is completed. Each passing vehicle is projected on the motion plane, and speed and dimensions are calculated by tracking it along the road. To test the accuracy of the proposed method, we constructed a ground-truth video dataset, by simultaneous capturing the road vehicles using RGB and speed cameras. Furthermore, to identify common vehicles, a dataset of vehicle images was collected. The proposed method evaluated on our dataset and the mean error of 1.15 km/h is achieved. Accordingly, it outperforms previous methods.

An Intelligent Measurement Method and System for Vehicle Passing Angles

“…However, different models of modified lifting plates were required to fit the widely varying vehicle requirements, meaning that it is still a contact measurement method that needs to be more convenient. Some automated methods are available to measure the geometric parameters of vehicles [8][9][10][11]. However, these methods cannot be directly applied to the measurement of passing parameters.…”

“…However, these methods cannot be directly applied to the measurement of passing parameters. The methods proposed in [8][9][10] employ various image segmentation techniques combined with vehicle morphology [8], neural networks [9], or prior knowledge [10] to extract vehicle contour boundaries and calculate their dimensions. Although these single-view methods are fast, they tend to have larger errors, and they struggle to capture information about the vehicle's bottom side using a single perspective.…”