PhD:
Pattern recognition & Template tracking Methodologies
ABSTRACT
In the present PhD thesis the problem of vehicle detection, location, classification and tracking in traffic video captures is examined and solved. We propose an operational system that empoyes two versions of appropriate software developed for this purpose. The first version concerns an algorithm constracted to detect vehicle line features on the gradient of traffic image sequence. The second version of this software employes a background subtraction algorithm appropriately modified to extract foreground information without apriori knowledge of the background (which apriori knowledge is often an empty road scene representing the initial background). Both algorithms have been developed and evaluated exploiting a list of road traffic captures, which have been collected from the two main Athenean avenues: Attiki odos and National Road E75.
A dataset of the aforementioned various road traffic captures has been created. The dataset’s road scenes captured by a plain commercial PAL camera under various lighting, angle view and traffic conditions. Appropriate software has been developed to manage those captures and they have been converted to frame sequences to increase flexibility of the system. The system requirements and experimental data for evaluation purposes for our two algorithms have been drawing from this dataset.
In our first approach, we have developed an algorithm capable of detecting and grouping line segments perpendicular to the traffic flow in binary images. According to our observations on the dataset, line segments perpendicular to the traffic flow is the distinctive feature that can safely separates vehicle figures from the background on image gradient (binary image). To achieve this goal we have employed a generalized Hough Transformation methodology to detect those line features. The detected line segments are grouped to form vehicle domains, whose domains trajectories have been processed using Kalman filter equations. The system was tested and found working satisfactory under normal traffic conditions.
In our second approach, we propose an innovating algorithm namely the Background Reconstruction. The Background Reconstruction algorithm is a heuristic that provides a periodically updated background and enhances the efficiency of the well-known background subtraction methodology in case of outdoor captures. This methodology guarantees a fresh instance of the actual background periodically, which is achieved by collecting scatter color information through a series of sequential images and assembling them to reconstruct the actual background. This process is applied to each pixel separately and the result is a color map of the actual image background.
The innovation of this study lies on the ability of the proposed algorithm to reconstruct the actual background color map without the need of any human intervention even in harsh traffic conditions, such as stop-and-go traffic flow. The background reconstruction algorithm demonstrated a rather robust performance in various operating conditions including unstable lighting, different view-angles and congestion.
We have designed two basic tests in order to evaluate the performance of the background reconstruction algorithm. In the first test we have chosen some specific pixels in a set of captures and we have separated the foreground from the background manually. In the second test we have chosen two popular algorithms, MOG and Codebook, to compare with. This comparison took place in a frame where we have first separated manually the foreground from the background and then we measured the separation performance of each algorithm.
The Reconstruction algorithm found outstanding in all tests. The capability of the algorithm to reconstruct the actual image background was confirmed according to the first test and it performed better than the other two algorithms in the second test. Further tests were processed by testing system performance in detecting vehicles and the background reconstruction capability in several captures of the dataset. The system managed to detect and track vehicles as well as to reconstruct the actual background. The processed traffic scenes as proof of algorithm’s reliability and robustness have been published on author web site at: http://users.ntua.gr/nmand/BGReconstruction.htm.
Publication:
A background subtraction algorithm for detecting and tracking vehicles, Expert Systems with Applications, vol. 38, pp. 1619-1631, 2011
ABSTRACT
An innovative system for detecting and extracting vehicles in traffic surveillance scenes is presented. This system involves locating moving objects present in complex road scenes by implementing an advanced background subtraction methodology. The innovation concerns a histogram-based filtering procedure, which collects scatter background information carried in a series of frames, at pixel level, generating reliable instances of the actual background. The proposed algorithm reconstructs a background instance on demand under any traffic conditions. The background reconstruction algorithm demonstrated a rather robust performance in various operating conditions including unstable lighting, different view-angles and congestion.