Traffic surveillance aims at detecting incidents and accidents to provide prompt and appropriate reaction. Tunnels deserve special safety concerns, since accidents in tunnels occur less often than outwards, but include a highly increased risk potential Automatized video surveillance is preferable, since humans suffer from operator fatigue on long term surveillance tasks. Tracking can be used to understand individual traffic participant's behaviour, allowing the extraction of stable properties such as vehicle velocity and trajectory in order to forecast incidents. The research presented in this paper shows how particle filtering can be used to enable vehicle tracking in traffic surveillance tasks. The presented method is qualitatively evaluated for a recorded real tunnel traffic data use case, while quantitative run-time performance analysis shows real-time capability on standard hardware. For evaluation we had access to a 8:19 minutes real world tunnel surveillance video sequence consisting of 12489 single images recorded with a frame rate of 25 Hz. This sequence contains 250 vehicles leading to a moderate traffic volume. Evaluating the video sequence using a preliminary detection module, 240 vehicles could be detected, but 5 of them were detected inaccurately resulting in initial states covering less than 50% of the vehicle visible in the image. Considering the other 235 vehicles correct tracking of 97.44% was achieved, where a vehicle is tracked correctly if more than 50% of the vehicle image is covered by the projected state in every time step until a maximum distance to the camera is reached. The maximum distance was chosen empirically, at a point where the perspective foreshortening becomes too big, resulting in tracking a vehicle for about 85 meters. Our implementation was tested on an Quad-Core Intel(R) Core(TM) i7 CPU 920 @ 2.67 GHz using OpenMP task parallelism over vehicle representations. Since the number of vehicle and hence the number of particles varies over the sequence, the mean value of a single vehicle tracked was computed over all images where at least one vehicle occurs, leading to an average computation time of 2.02 ms tracking time per frame. The standard deviation of 1.12 ms appears to the costly initialization phase and the influence of interpolating image regions of larger vehicles during scaling.
Vehicle tracking using 3D particle filter in tunnel surveillance and incident detection
2013
10 Seiten, Bilder, 12 Quellen
Conference paper
English
Incident Detection Using Vehicle-Based and Fixed-Location Surveillance
British Library Online Contents | 1997
|Incident Detection Using Vehicle-Based and Fixed-Location Surveillance
Online Contents | 1997
|British Library Online Contents | 2012
|Multiple Model Particle Filter for Traffic Estimation and Incident Detection
Online Contents | 2016
|