From: Subject: Tracking by Cluster Analysis: Results on the PETS2000 Image Sequences Date: Thu, 4 Nov 2010 13:39:06 +0100 MIME-Version: 1.0 Content-Type: text/html; charset="utf-8" Content-Transfer-Encoding: quoted-printable Content-Location: http://www.diku.dk/OLD/publikationer/tekniske.rapporter/rapporter/01-07/ X-MimeOLE: Produced By Microsoft MimeOLE V6.00.2900.5994 =EF=BB=BF Tracking by Cluster Analysis: Results on the PETS2000 = Image Sequences

Tracking by Cluster Analysis: Results on the PETS2000 Image = Sequences

DIKU Technical Report no. = 2001/07=20
ISSN: 0107-8283



This technical report contains two videos obtained with the = Cluster=20 Tracker described in [1]. This = web page=20 will give a short introduction to the Cluster Tracker and a short = description of=20 the videos: the reader is referred to [1] for a = full=20 description of the tracker.

Many tracking systems include the four stages of image differencing,=20 thresholding of the difference image, morphological filtering, and=20 connected-component labelling. These stages are used to identify = distinct=20 targets and attribute each image pixel to one of the targets. Further = processing=20 stages use this information for detection of target features, Kalman = filtering,=20 etc. The approach can be quite effective, but thresholding and = morphological=20 operators involve information loss: this loss is the result of assigning = each=20 pixel unambiguously either to the background or to one (and only one) = target,=20 ignoring the uncertainty of these assignments.

For most purposes, this attempt is unnecessary: the output that is = required=20 from a tracker is information on how many targets are present and the=20 approximate location and size of the targets. The Cluster Tracker = developed by=20 the author obtains this information by cluster analysis of pixels. The=20 probabilities that a pixel belongs to the background or to one of the = targets=20 are estimated from the location and grey-level value of the pixel. = Cluster=20 parameters are optimized by the EM algorithm and the number of clusters = is=20 determined by increasing or decreasing the numbers of clusters by some = simple=20 statistical tests.

The algorithm was tested on the PETS 2000 image sequences. More = information=20 on the PETS workshops is available from http://visualsurveillance.org/

The image sequences were downsampled 2x2 times in image space and 5 = times in=20 time, i.e. the tracker used 5 frames/second. The specific implementation = of the=20 Cluster Tracker used to obtain the results presented in the videos is = fully=20 described in [1], = including all=20 parameter settings.=20

The original image sequences have a rate of 25 frames/second, which = means=20 that e.g. frame 1200 corresponds to 48 seconds: this should be kept in = mind when=20 interpreting the frame numbers given below.

Cluster=20 tracking in the PETS2000 Test Sequence

The three vehicles and three people visible in the test sequence were = detected and tracked until they left the field of view, or until the end = of the=20 image sequence. Two birds appearing in the upper right corner were also=20 detected. The second bird was merged into a specular reflection from the = building which appeared at about the same time. The ellipsoids represent = two=20 "standard deviations" of the Gaussian clusters (see [1] for = more=20 details): they can be seen to fit quite well to the actual targets; the = fit is=20 even better if the shadows of the targets on the ground are taken into = account.=20

The pedestrian entering the field of view from the top is not = detected=20 immediately, because she is initially too far from the camera. Some more = sensitive detection technique might result in earlier detection. On the = other=20 hand, some false alarms are caused by specular reflections, and a lower=20 threshold would result in a higher rate of such false alarms.

Some additional techniques that can limit the number of false alarms = (not=20 used in the videos shown here) were described by the author in the = proceedings=20 of the PETS 2001 and PETS 2002 workshops (see references [2] and [3] = below).=20

The main problem with the tracker becomes evident in frame 1200: the=20 pedestrian who has been tracked from frame 350 is merged, first into the = parked=20 car, then into the same cluster as the second passenger leaving the car = (frame=20 1255), before acquiring again her own cluster. Of course, at this point = she is=20 no longer recognized as the same target that merged into the car.

Cluster=20 tracking in the PETS2000 Training Sequence

In the PETS2000 training sequence, all vehicles and pedestrians were = detected=20 and tracked successfully, except for a temporary ``merging'' of two = pedestrians=20 into a single cluster (frame 1510), which results in the swapping of = their=20 identities. When these two pedestrians reach their car, they merge into = a=20 specular reflection from the car (frame 1600); when the car starts to = move, it=20 carries along the identity of one of the pedestrians. In this case, the=20 pedestrian was inside the car, but it could have been otherwise.

As in the test sequence, there were a few false alarms, due to = variable=20 specular reflections from cars and from the building in the background. = One car,=20 which was included in the reference image, leaves behind a ``ghost'' = (the=20 cluster at the top left of frame 4075 and of following frames) when it = moves out=20 of its parking place: the problem is not due to the tracker, but to the=20 initialization of the reference image. Note also from frame 4075 that a=20 pedestrian can occasionally be split into two clusters; however, the two = clusters have a common motion and (at least in this image sequence) = merge again=20 at a different phase of the walking cycle.

As in the case of the test sequence, the additional techniques = described in=20 [2] and [3] might be effective in reducing false alarms. However, this = technical=20 report only describes results with the basic Cluster Tracker.

On the positive side, the algorithm can cope pretty well with = reversals of=20 the driving direction, as can be seen e.g. in frame 4305.

References

  1. A.E.C. Pece (2002)=20
    Generative-Mod= el-Based=20 Tracking by Cluster Analysis of Image Differences.=20
    Robotics and Autonomous Systems vol. 39 (3-4), pp. = 181-194.
  2. A.E.C. Pece (2001)=20
    Tracking of=20 Non-Gaussian Clusters in the PETS2001 Image Sequences.=20
    Proceedings of the 2nd IEEE Int. Workshop on Performance = Evaluation in=20 Tracking and Surveillance:=20
    PETS2001, Kauai (Hawaii), 9 Dec. 2001.
  3. A.E.C. Pece (2002)=20
    From = Cluster=20 Tracking to People Counting=20
    Proceedings of the 3rd IEEE Int. Workshop on Performance = Evaluation in=20 Tracking and Surveillance:=20
    PETS 2002, Copenhagen, 1 June 2002, pp. 9-17.=20