DTA: Deployment and Tracking Algorithm in Wireless ...

DTA: Deployment and Tracking Algorithm in Wireless Multimedia Sensor Networks Ibtissem Boulanouar, Stéphane Lohier, Abderrezak Rachedi, Gilles Roussel Université Paris-Est - Gaspard Monge Computer Science Laboratory (LIGM – UMR 8049) 75454 Champs sur Marne, France {FirstName. LastName}@univ-mlv.fr

Abstract—In this paper we address the problem of node deployment and target tracking in Wireless Multimedia Sensor Networks. We define Mobile Object Tracking as a two stage process: 1) detecting the presence of the object in the monitoring area; and 2) locating it at each stage of its progression in the area of interest. We propose DTA: Deployment and Tracking Algorithm, a distributed protocol that runs on Heterogeneous WSN (HWSN), composed of two types of sensors: Motion Sensors (MSs) and Camera Sensors (CSs). DTA starts by a deployment strategy for both sensors type based on critical sub-areas. Then, the tracking algorithm begins. The MSs deal with the detection phase and activate the CSs which are the most likely to detect the mobile target. The heterogeneous network and the deployment strategy for this network represent the main contribution of this paper. We carry out simulations to evaluate the efficiency of our proposed solution in terms of tracking accuracy, energy consumption and number of exchanged messages. The obtained results are compared to two existing solutions: 1) BASIC solution where only CSs are deployed in active mode to monitor the area of interest; and 2) OCNS (Optimal Camera Node Selection), a cluster-based approach with CS election. We observe that the deployment strategy has a positive impact on target tracking. We also notice that DTA significantly reduces the tracking cost in term of energy consumption (up to 35.87% of energy saved compared to other algorithms). It also performs tracking with better accuracy (up to 37.5% on the tracking accuracy).

I.

INTRODUCTION

Advances in communication technologies have enabled the development of small sensors with different sensing, processing and communication capabilities. A Wireless Sensor Network (WSN) consists of a set of Wireless Sensors which can harvest, process and share data in an area of interest. Nowadays, sensors allow handling more complex data such as multimedia flows. Thus, we observe the emergence of Wireless Multimedia Sensor Networks (WMSN). These WMSN offer a wide range of applications such as intrusion detection, target tracking, habitat and health care monitoring. In addition to the common characteristics shared with WSN, the WMSN have special features: high bandwidth demand, specific QoS requirements, sector sensing range, etc. Moreover, Wireless Camera Sensors (CSs) unlike classical video systems can be used in both indoor and outdoor environments, where energy and network infrastructure are not available and where no human intervention is possible. They offer a wider panel of applications whether for environmental, industrial or military monitoring [1].

1

In this paper, we focus on deployment strategy for object tracking in WMSN and more particularly on noncommunicating object tracking like human beings, animals, etc. Mobile target tracking consists in retrieving the target’s successive coordinates during its moving through the monitoring area. This topic has been well studied in classical and omnidirectional WSN, particularly in the case of communicating object tracking. With the recent advances in both miniaturization and performances of CMOS technologies and embedded video processing technique, tracking mobile object using video sensor wireless networks has entailed a lot of interest from the sensor networking community. For that purpose, we propose DTA: Deployment and Tracking Algorithm for Multimedia Sensor Networks. Optimal sensor deployment is a challenging issue and depends on the knowledge of surrounding environment. DTA runs on Heterogeneous Wireless Sensor Network (HWSN) composed of two types of sensors. Motion Sensors (MS) are infrared detectors and Camera Sensors (CSs). The advantage of using this type of network is to improve the network performance and lifetime by sharing the tasks between the different kinds of sensors: high cost energy tasks can be achieved sporadically by powerful sensors while low-cost energy tasks can be done by constrained sensors. DTA has two main phases: deployment and target tracking. A deployment strategy for both kinds of sensors is proposed in the first phase. We introduce the notion of critical sub-area: in real indoor environment, some areas are more important to cover than other (entrance/exit, corridor, etc.). Our proposed deployment technique begins with MS placement. We divide the Area of Interest (AoI) in cells grid and place the MSs at the center of each cell. Afterward, we place the CSs, with random orientation, at the corner of every critical sub-area. Finally, every CS calculates its most beneficial orientation using local information such as neighbors and critical sub-areas. In target tracking phase, we propose an energy-aware and collaborative tracking algorithm. This tracking algorithm is based on cooperative communication between the two types of sensors. The MSs that handle the detection consume less energy than CSs which allows to keep them actives in order to monitor the AoI [2]. CSs are in charge of localization. When a MS detects a mobile object, it activates only the CS that can localize it. The performance of the tracking algorithm is closely related to the deployment strategy. Indeed, an efficient deployment strategy ensures a maximal coverage of the area of interest which increases the tracking algorithm performances. DTA is based on our previous proposed work called EAOT [3]. DTA brings significant improvement as it gives more accurate tracking results and allows saving more energy that the existent tracking algorithms which run on homogenous WMSN. The main characteristics of the proposed DTA are: 1) detection of the target when it enters the monitoring area; 2) performing the tracking in an energy-aware manner; 3) object localization based on three possibilities: real coordinates, a visual observation of the target, and as worst case approximate coordinates . The reminder of this paper is organized as follows: Related Work is reviewed in Section II. In Section III we describe our proposed deployment strategy and tracking algorithm DTA. In Section IV, we discuss the performance evaluation. Finally, the paper concludes with Section V.

2

II. RELATED WORK A. Deployment Strategy Over the last years, many sensor deployment methodologies have been studied. They can be categorized in two main policies: Random and Planned [4]. The choice of deployment policy depends on the region topology, sensors features and application objective. Random deployment has the advantage of being simple and fast, this is especially suitable when nothing is previously known about the AoI. It is used in some outdoor applications such as rescue in disaster area and military applications. In planned deployment, the sensors are placed following a predefined method. It is often used in indoor applications and necessary when nodes performances are affected by their location. We denote two kinds of sensors: omnidirectional and directional. The most used solution for omnidirectional sensors is the grid approach where the Area of Interest is divided in grid of cells and the sensors are placed at the intersection of these cells. The cell size is generally related to the node sensing range. The cells can be either squares [5] or triangles [6]. In [7] the AoI is decomposed in Voronoi diagram. The objective is to determine the minimum distance to any nodes that any mobile object traveling through the monitoring region must have, even if the mobile object tries to avoid the nodes. Using an optimal polynomial time algorithm the worst case breach coverage is computed. This solution allows finding not covered area. This information is useful in some applications such as intrusion detection and target tracking. In [8] the authors propose a sensor deployment strategy that considers environmental features and the number of sensors. The proposed strategy consists of four steps:

the installation of a sensor network, the analysis of

environmental features, the sensor deployment, and the selection of the relay sensor. The deployment is performed considering network features as well as area factor. For the directional sensors, specific solutions already exist. Given a set of point-locations, the work presented in [9] attempts to reduce the network cost by minimizing the number of deployed directional nodes and adjusting nodes parameters: Field of View (FoV), orientation and sensing range. For that purpose the authors use an integer linear programming solution. The solution proposed in [10] is the most popular of them; it is based on a Virtual Force Algorithm (VFA). This algorithm calculates the virtual force applied by neighborhood on the centroid of each directional sensor view field, to determine its new position and direction. In [11], FVPTR algorithm is proposed. FVPTR for Fuzzy Image Recognition and Virtual-Potential-field-based paired Tangent point Repulsion has the same working principal as VFA but for a large scale network. In [12], the authors proposed a new algorithm called PSO, Particle Swarm Optimization algorithm to solve the sensor deployment problem. PSO is an iterative and meta-heuristic optimization algorithm. It imitates the swarm behavior of some organisms like fish or birds. Using this algorithm, the most beneficial position and orientation of each node in the network is found. In these works, the concept of critical sub-areas is neglected. In this paper we introduce this concept; we propose a deployment strategy for both kind of sensors and we study its impact in tracking application.

3

B. Target tracking Mobile Object Tracking is a challenging issue in WSN research field. It consists in detecting and locating the mobile object throughout its moving stages in AoI. The existing researches about target tracking in WSN can be classified in three main categories: predictive-based, cluster-based, and structure-less. The first class depends on tracking relaying strategy while the lasts depend on network architecture. In the predictive based class, predictive models are used to proactively predict the future location of the mobile object. Then, the selected nodes are activated along the target path for energy saving. In [13] the authors proposed an autoregressive model to model and predict the target movement in cellular networks. Although this algorithm is effective, it requires a sampling phase to initialize the parameter, which has a non-negligible cost. With the same objective, in [14], a Gauss-Markov model is used to model target behavior, then, predicts its future displacement. In [15] the authors proposed a dual-prediction mechanism to achieve energy savings by keeping most of nodes in sleep mode. However, this solution is not reliable in applications like military tracking as it cannot ensure a sufficient accuracy of the trajectory predicted. Cluster-based technique [16, 17], is the most popular one; a set of sensors is selected at each step of the tracking process. A cluster is composed of cluster head and cluster members. The clusters can be formed statically at the network deployment stage or dynamically, based on predefined criteria such as position, processing power, residual energy, etc. In [18] the authors proposed an energy saving self-organizing protocol called LESOP for mobile target tracking in WSN. A lightweight and efficient tracking algorithm is proposed to handle the tradeoff between the tracking accuracy and the energy consumption. In [19], the authors studied the node selection in WMSN, where a visual observation of the mobile target is available. They proposed OCNS for Optimal Camera Network Sensor, a cluster based and two phases cooperative algorithm: target detection and target localization. The main objective of this solution is to maintain the desired subset of Camera Sensors in active mode while the others remain in sleep mode. When a Camera Sensor detects the moving object, it broadcasts its own coordinates to all the nodes that are within its transmission range. Each of them calculates the probability of detecting the mobile target. If this probability exceeds a defined threshold, the node activates its camera to perform the localization. A wake up algorithm is also merging with this solution to keep the desired number of nodes in active mode. SensEye [2] is the first work that introduces the concept of heterogeneous wireless network. The authors propose three-tier camera sensors; every tier supports a specific task. The first tier assumes mobile target detection and localization while the second one performs target recognition and the last tier assumes target tracking. SensEye is dedicated only to indoor applications where energy is available which considerably limits the types of applications where we can use it. In structure-less approach, no network architecture is defined. The tracking is performed in reactively manner at each stage of target evolution inside the area of interest. Our proposed solution DTA is considered as structure-less based solution. Indeed, no predefined network architecture is set up. DTA, unlike existing works is a global solution which includes both deployment strategy and tracking algorithm. The deployment strategy is proposed to enhance the performance of the tracking algorithm. To our best knowledge, DTA is the first work that proposed the deployment for Heterogeneous Wireless Sensor Networks.

4

III. SOLUTION OVERVIEW In this section, we present the proposed deployment strategy and the protocol called DTA: Deployment and Tracking Algorithm. Its main objectives are as follows: 1) having a visual observation of the mobile target, which permits its identification; 2) having target coordinates to use them for specific applications like intrusion detection; 3) saving energy to increase the network’s lifetime. A. Preliminaries and Assumptions In this subsection we give some definitions and assumptions. Definition 1: The Motion Sensors (MS) are scalar sensors fitted with infrared motion detectors. Each of them has a circular Field of Detection (FoD) with radius RD (Figure 1.A). The Camera Sensors (CS) are multimedia sensors equipped with video cameras. Each of them has a directional Field of View (FoV) defined by a cone with radius RV and angle α (Figure 1.B). For example the imote2 sensor proposed by MEMSIC [20] (formerly CrossBow) is equipped with an infrared detector and a low-resolution (640x480) video camera which can be independently activated. An individual mobile object is expected to cross the AoI by taking a random trajectory. As defined above, our HWSN consists of a set of MS and CS placed according to our proposed deployment strategy in monitoring region in order to detect, observe and locate a mobile target. All MSs have the same circular FoD and all the CSs have the same FoV. Every MS is aware of its location and every CS is aware of its location and orientation via one of the existing solutions [21]. Each sensor can communicate with one other, independently of its type, as long as they share the same transmission range (RT). Initially all the cameras are supposed to be in sleep mode and the Motion Detectors are always in active mode.

(B)

(A)

Figure 1: FoD and FoV

Definition 2: The term approximate coordinates, (Xt, Yt) of the target is introduced. They represent the target coordinates calculated by the MSs when no CS is available. We calculate them as follow: X ! = P!" Y! = P!"

! !!! !!

! ! !!! !!

… (1)

!

Where PMS is the probability that the mobile object is detected by MS. (Xm, Ym) are the coordinates of MS. N is the total number of MSs that detect the target at the same time in a given area. As detailed in [22], the value of PMS is 5

closely related to the distance  𝑑 X, Y , X ! Y! between the target and the sensor. Where (X, Y) represent the real target coordinates. We can express PMS as follow: PMS=

𝑒

1111111, 𝑖𝑓  𝑑((𝑥, 𝑦), (𝑖, 𝑗)) ≤ 𝑅! , 𝑖𝑓  𝑑((𝑥, 𝑦), (𝑖, 𝑗)) > 𝑅!

!!  !((!,!),(!,!))

β define the physical characteristics of the MS. We observe that PMS decreases exponentially while 𝑑 𝑥, 𝑦 , 𝑖, 𝑗 increases. Definition 3: PCS is the important new concept introduced in this work. It represents the probability that the mobile object is detected by CS. It depends on three parameters: 1) the number of MSs in CS’s transmission range that detects the target; 2) the distance between CS and these MSs and 3) the orientation of CS. PCS is obtained as follows: !

𝑃!" = 1 −

1 − 𝑃!"# … (2) !!!

Proof:

             𝑃!" = (𝑃!"! + 𝑃!"! + 𝑃!"! +   … + 𝑃!"# )     = 𝑃!"! ×𝑃!"! ×𝑃!"! …×𝑃!"#  = 1 − 𝑃!"! ×𝑃!"! ×𝑃!"! …×𝑃!"# =1−

! !!!

1 − 𝑃!"#

Hence, we can express the probability PCSi that the mobile object is detected by CS depending only on one MS by equation (3).

Figure 2: Initial deployment strategy 𝑃!"# =

𝐴!"# … (3) 𝐹𝑜𝐷

Aint is the intersection area between FoD and FoV (see black dashed line in Figure 2). It is calculated on the basis of the distance between CS and MS: the value of Aint decreases with the distance. The distance is calculated on the basis of the MS and CS coordinates, known by each of them using the solution described in [21].

6

B. Deployment Optimization In this sub-section we introduce our proposed deployment strategy. For the two types of sensors, we face two main constraints: the MS placement, the CS placement and orientation. 1) Initial deployment scheme: We firstly consider the deployment of MSs. We propose to divide the AoI in grid of cells where each cell has the size of MS detection radius. Unlike the solution described in [5], we place MS at the center of cells. Thus, we avoid overlaps. The number of MS is equivalent to the number of cells. As we say in previous section, we introduce the notion of critical sub-areas. Therefore, the total number of CSs is calculated according to their number. For convenience, we suppose that these critical sub-areas have square shapes. So, the number of deployed CS is equivalent to the number of critical sub-areas corners. When the total number of CS is obtained, we place them with random orientation at the corner of each critical sub-area. Then we try to optimize the orientation of each CS by considering local information like neighbor’s orientation and critical sub-areas. Because of the higher cost of the CSs, we can’t deploy enough to cover the whole area. In addition, because of the sector sensing range, we need more CSs than MSs to cover all the area of interest. In our work, we consider that the critical sub-areas are more important to cover than the rest of the area. For that purpose, we place initially the CSs at each corner of the critical sub-area. The coverage of critical sub-area depends on the size of CS’s FoV and the size of critical sub-area. In this work, we suppose that we cover the critical sub-area with four CSs. An illustrative example of initial deployment strategy is shown in Figure 2. The AoI is divided into grid of cells. The Motion Sensors are placed at the center of these cells. An example of critical sub-area and how we place the CSs are shown. 2)Camera sensor orienting scheme: We propose a distributed mechanism to find the most beneficial orientation of each CS according to critical subareas. 2.a. Collecting information phase: in this first phase each node collects useful information about its surrounding environment.

Figure 3: 2D FoV

7

Neighbors list: Every CS has to construct a Neighbors-list. For that purpose each of them broadcasts a HELLO message containing its ID and its own coordinates to all neighbors. We consider neighboring CS as the CSs which shared a common FoV. This condition is realized when the Euclidian distance between them