The video camera systems that existed 20 years ago were displaced within a very short time for classification and volumetric tasks, since the distance information of the laser scanners guarantees in principle a very simple, robust and non-computationally intensive analysis. However, laser scanners are now reaching their technical limits, in particular due to the finite scan rate, which (especially for fast vehicles) only allows a very limited spatial resolution. This effect is exacerbated by adverse weather conditions.
In the project "Contactless axle detection for traffic monitoring", the project partners were able to convincingly demonstrate that vision-based axle detection in flowing road traffic significantly outperforms the detection accuracy of laser scan systems. Therefore, it is becoming apparent that video camera systems are reclaiming the market because these systems, in principle, provide depth information in addition to gray scale or color information by fusing data from multiple cameras, and do so at full video rate and resolution.
Therefore, the aim of this project is to investigate a method for vision-based counting and classification of vehicles. The main USP is to achieve the TLS 8+1 A1 standard for the first time for a vision-based product on the market and thus to surpass the performance of current laser scan-based systems.
For this purpose, the immense scientific-technological progress of the last years in computer vision shall be used in the project, both in the hardware development (camera and embedded computer platforms) and the algorithms for machine learning, especially for the classification with deep learning methods.