A SMART CAMERA APPROACH TO REAL-TIME TRACKING

Tracking applications using distributed sensor networks are emerging today, both in the field of surveillance (airports, train stations, museums, public spots) and industrial vision (visual servoing, factory automation). Traditional centralized approaches

A SMART CAMERA APPROACH TO REAL-TIME TRACKINGSven Fleck, Sven Lanwer, Wolfgang StraßerWSI/GRIS, University of T¨ bingen u Sand 14, 72076 T¨ bingen, Germany u phone: +(49) 7071 2970435, fax: + (49) 7071 295466, email: { eck,lanwer,strasser}@gris.uni-tuebingen.de web: www.gris.uni-tuebingen.deABSTRACT Tracking applications using distributed sensor networks are emerging today, both in the eld of surveillance (airports, train stations, museums, public spots) and industrial vision (visual servoing, factory automation). Traditional centralized approaches offer several drawbacks, due to limited communication bandwidth, computational requirements and thus also limited spatial camera resolution and framerate. In this paper we present a network-enabled Smart Camera for probabilistic tracking. It is capable of tracking objects in real-time and offers a very bandwidth-conservative approach, as it only transmits the tracking results which are on a higher level of abstraction. 1. INTRODUCTION Today’s computer vision systems typically see cameras only as simple sensors. The processing is performed after transmitting the complete raw sensor stream via a costly and often distance-limited connection to a centralized processing unit (PC). We think it is more natural to also physically embody the processing in the camera itself: what algorithmically belongs to the camera is also physically performed in the camera. The idea is to compute the information where it becomes available – directly at the sensor – and transmit only results that are on a higher level of abstraction. This follows the emerging trend of self contained and networking capable Smart Cameras. We present a rst prototype of a network-enabled Smart Camera capable of probabilistic object tracking in real-time. Tracking plays a central role for many applications including robotics (visual servoing, RoboCup), surveillance (person tracking) and also human-machine interface, motion capture, augmented reality and 3DTV. Particle lters have become a major way of tracking objects [1, 2, 3]. Utilized visual cues include shape [3] and color [4, 5, 6, 7] or a fusion of cues [8, 9]. The particle lter algorithm is described in section 2. We use a color histogram based approach adapted to the special needs of our hardware target. Our Smart Camera tracking architecture is described in section 3. Afterwards, we discuss various bene ts of our approach and show experimental results in section 4 before we conclude this paper. 2. PARTICLE FILTER Particle Filters can handle multiple hypotheses and nonlinear systems. Following the notation of Isard and Blake [3], we de ne Zt as representing all observations {z1 , ..., zt } up to time t, while Xt describes the state vector at time t with dimension k. Particle Filtering is based on the Bayes rule to obtain a posterior p(Xt |Zt ) at each time-step using all available information: p(Xt |Zt ) = p(zt |Xt )p(Xt |Zt 1 ) p(zt ) (1)weight p , with(i) stdenoted by Together they form the sample set (i) St = {st |i = 1..N}. Fig. 1 shows the principal operation of a Particle Filter with 8 particles, whereas its steps are outlined below.N p (i) = 1. i=1 (i) (i) = (Xt , p t ).Thus, the i-th sample at time t isp X t 1 Zt1[Choose] & [Prediction] Deterministic Prediction through Motion Modelp Xt Xt[Diffusion]1[Measurement]p zt X tp X t Ztp zt X t?p X Xtt 1p X t 1 Zt1dX t1Figure 1: Particle Filter iteration loop Choose Samples Step: First, a cumulative histogram of all samples’ weights is computed. Then, according to each par(i) ticle’s weight p t 1 , its number of successors is determined according to its relative probability in this cumulative histogram. Prediction Step: In the prediction step, the new state Xt is computed: p(Xt |Zt 1 ) = p(Xt |Xt 1 )p(Xt 1 |Zt 1 )dXt 1 (2)Different motion models are possible to implement p(Xt |Xt 1 ). We use three simple motion models (whereas the speci cation of how many samples belong to each model can be parameterized): a random position model, a zero velocity model and a constant velocity model (Xt = AXt 1 + wt 1 ), each enriched with a Gaussian diffusion wt 1 to spread the samples and to allow for target moves differing from each motion model. Our (i) (i) state has the form Xt = (x, y, vx , vy )t . Measurement Step In the measurement step, the new state Xt is weighted according to the new measurement zt (i.e., according to the new camera image). p(Xt |Zt ) = p(zt |Xt )p(Xt |Zt 1 ) (3)whereas this equation is evaluated recursively as described below. The fundamental idea of Particle Filtering is to approximate the probability density function (pdf) over Xt by a weighted sample set St . Each sample s consists of the state vector X and aThe measurement step (3) complements the prediction step (2). Together they form the Bayes formulation (1).

Tracking applications using distributed sensor networks are emerging today, both in the field of surveillance (airports, train stations, museums, public spots) and industrial vision (visual servoing, factory automation). Traditional centralized approaches

2.1 Color Histogram based Particle Filter Measurement Step in context of Color Distributions As already mentioned, we use a particle lter on color histograms. This offers rotation invariant performance and robustness against partial occlusions and non-rigidity. In contrast to using standard RGB space, we use a HSV color model: A 2D Hue-Saturation histogram (HS) in conjunction with a 1D Value (V ) histogram is designed as representation space for (target) appearance. This induces the following specializations of the abstract measurement step described above. From Patch to Histogram Each …… 此处隐藏：18561字，全部文档内容请下载后查看。喜欢就下载吧 ……