Lecture consists of a short explanation of the LBP feature extraction and classification algorithms, architecture and experimental results of the proposed system for vehicle detection and tracking from the video footage. Dataset used for the classifier learning process is obtained from the video footage made by Faculty of Transport and Traffic Sciences and video footage obtained from Toyota Motor Europe website. Testing of the system is made on a video footage of Zagreb bypass.

On Tuesday, September 16th, joint VISTA and CCVW workshop has been held at Faculty of Eletrical Engineering and Computing (UNIZG-FER). At the workshop Kristian Kovačić has held the presentation “Real Time Vehicle Trajectory Estimation on Multiple Lanes“. Published article can be downloaded here.

A research seminar in scope of VISTA project with title “Real-time vehicle detection and tracking on multiple lanes using one camera” will take place on 15. April, 2014. at 14:15. Seminar will be held at Faculty of Electrical Engineering and Computing (UNIZG-FER) in TCR hall.

Abstract:

Development of computing power and cheap video cameras enabled today’s traffic management systems to include more cameras and computer vision applications for transportation system monitoring and control. Combined with image processing algorithms cameras are used as sensors to measure road traffic parameters like flow, origin-destination matrices, classify vehicles, etc. Typical commercial computer vision based traffic monitoring systems use one camera per lane to ensure accurate and robust traffic parameters measurement. This presents a drawback since many cameras are needed for roads with multiple lanes making such systems expensive. In this presentation main accent will be on methods that allow real-time multiple vehicle detection and tracking on multiple lanes using one camera only. Vehicle detection is based on separating foreground (dynamic) and background (static) segments in an image. Background model is determined comparing a large number of preprocessed images with a background image and updating it accordingly. All dynamic pixels in an image are grouped into clusters by checking their vicinity with other dynamic pixels. Tracking algorithm computes a weight factor for each cluster based on a cluster area, cluster overlapping area and distance between multiple clusters. Several past vehicle poses and pose changes are used to enhance the vehicle trajectory estimation. To ensure real-time capabilities, image processing algorithm computation distribution between CPU and GPU is applied. The CPU executes parts related to pixel clustering and object tracking while GPU executes parts that can be parallelized for computation (image preprocessing, creation of a background model, moving object detection). Described system is tested using real traffic video footage obtained from Croatian highways.