Wireless sensor networks have been the subject of intense research in recent years. Sensor nodes are used in wide range of applications such as security, military, and environmental monitoring. One of the most interesting applications in wireless sensor networks is target tracking, which mainly consists in detecting and monitoring the motion of mobile targets. In this paper, we present a comprehensive survey of target tracking approaches. We then analyze them according to several metrics. We also discuss some of the challenges that influence the performance of tracking schemes. In the end, we conduct detailed analysis and comparison between these algorithms and we conclude with some future directions.
Recent advances in MEMS technology have given rise to low power, cost, and multifunctional sensor nodes [
Target tracking (shown in Figure
Scenario of target tracking in wireless sensor networks.
Target tracking can be performed using a single node or through the collaboration between different sensors. However, using a single node may result in power loss and induces heavy computation burden on that node, while using multiple sensors gives better results in terms of accuracy and energy saving, due to the cooperation between nodes.
Various taxonomies of target tracking algorithms were proposed in the literature. So, there is no standardized or predefined classification. Some works have studied tracking algorithms according to security aspect [
The rest of the paper is organized as follows. Section
Many challenges can affect the target tracking quality in wireless sensor networks: Node failure: nodes in WSNs are prone to failures due to battery exhaustion, physical disasters, hardware failures, external attacks, and so on. Consequently, proposed target tracking protocols should cope with these challenges. Target missing and recovery: prediction errors, obstacles, sudden changes in target trajectory, and speed cause loss of target. To tackle this challenge, robust tracking algorithms should be proposed to decrease the probability of missing targets. Moreover, recovery mechanisms have to be considered in case the target was lost. Coverage and connectivity: there is a mutual relationship between coverage and target tracking. High coverage results in high tracking accuracy. However, the performances of target tracking algorithms are degraded in the case of sparse networks, or in the presence of coverage holes. Data aggregation: the aggregation mechanism [ Tracking latency: the execution of target tracking algorithms must be performed rapidly while preserving positioning accuracy. When the operation of tracking takes too long, the moving node may change its location. Energy consumption: since sensor nodes run on batteries which are, in most cases, nonrechargeable, energy efficiency is a critical issue in wireless sensor networks especially in sensitive applications like target tracking.
Target tracking algorithms were studied in the literature from several angles, so there is no standardized classification. Target tracking can be classified according to different aspects: security, energy efficiency, network structure, accuracy, mobility of the target, fault tolerance, and so on.
In this section we will discuss the metrics according to which we will evaluate several recent algorithms of target tracking. As shown in Figure
Target tracking possible metrics.
In [
Meanwhile in [
We have classified the network structure into three categories: tree structure [
Static clustering: in static clustering, clusters are formed at the time of network deployment and remain unchanged until the end of network lifetime. Despite its simplicity, the static clustering [ Dynamic clustering: the clusters are formed dynamically as the target moves. The use of dynamic clustering has many benefits. This is due to its flexibility; that is, new clusters are established as the need arises. Furthermore, only one cluster is activated when the mobile target passes by, which implies the mitigation of data redundancy and interference issues.
Prediction methods are used to predict the future position of the mobile object. Only sensor nodes located near this position are turned on to detect the target, while the other nodes remain in sleep mode to conserve energy. Prediction techniques are often integrated with face, cluster, or tree structure. Different techniques are used to predict the next position of the target such as kinematics, Kalman filter (KF) [
Target tracking approaches are divided into two types: single and multiple target tracking. Single target: in general, tracking a single object is energy efficient and consumes less power. This is because only a low traffic load is generated in the network while chasing the target. Multiple targets: tracking multiple objects is a very challenging task, especially when the number of targets arises. It becomes more complex because of the different directions and speed variations of the tracked objects. In cluttered environment, and in the presence of multiple targets, a sensor node can obtain more than one measurement. It is difficult to know which observation belongs to which target. This uncertainty in measurements results in data association problem.
Most of the proposed approaches in target tracking aim to track individual (called also discrete) objects such as vehicles, people, and animals. However, few research efforts have been done on tracking continuous objects such as forest fires, gas leakage, biochemical material diffusion, and oil spills. In contrast to discrete objects, the continuous ones occupy a large area, tend to diffuse, change in shape, increase in size, and even split into smaller objects.
Sensor node can be either binary or ordinary. Binary sensors detect the presence or absence of the target in its sensing range by generating one-bit information (0 or 1). They are low-cost and consume less energy. By using binary outputs, the size of data sent to the base station is reduced.
Prediction algorithms may suffer from localization errors. This is why a recovery mechanism should be included while designing target tracking approaches in order to relocate the missing object.
Target tracking can also be studied from the network architecture aspect. We can distinguish two types of architecture.
In this section, we survey the state of the art of some target tracking protocols in wireless sensor networks, and then we will evaluate them according to the metrics defined in Figure
In [
The static head algorithm performs worse than the adaptive head in terms of energy efficiency. This goes back to the fact that a large number of control messages are exchanged during each cluster formation. Whenever a target is detected, slave nodes and neighboring cluster heads send their distance readings to the master node which congest the network.
The rate of tracking errors in the adaptive head algorithm increases as the target speed increases. This is due to error accumulations produced when estimating the position of the target as well as the noise added to TOA [
One of the biggest issues in static clustering is the boundary problem. Sensor nodes in different clusters are not able to collaborate or exchange information with each other, which leads to a boundary problem when the mobile target moves along the boundaries of these clusters. To solve this issue, authors in [
Boundary problem.
The dynamic clustering employed in HCTT approach induces too much overhead while forming, maintaining, and dismissing the clusters. These operations will consume a lot of energy and lead to the premature death of sensor nodes.
In [
The cluster lifetime.
The dynamic clustering in this approach brings extra overhead. A reclustering process is triggered whenever the target changes its location. Furthermore, each time a new cluster head is elected, it broadcasts advertisement messages to the other sensor nodes to announce its election. As a consequence, these advertisements reduce drastically the network lifetime and its performances.
Upon its movement, the target may enter into a hole [
Authors in [
This approach consumes more energy, since, in each measurement period, sensor nodes transmit a large amount of packets towards their cluster head.
A variational filtering (VF) approach was proposed in [
Binary sensors do not provide accurate estimation of the target’s location, direction, or velocity. They only generate one-bit information regarding the presence or absence of the target.
Authors in [
FOTP approach.
In FOTP approach, when the target gets lost and could not be detected in a predefined time
Hsu et al. have proposed a Prediction-Based Optimistic Object Tracking (POOT) strategy [
CORS algorithm saves more power in comparison with TORS. This is because CORS activates only faces where the missed target most possibly resides. Such behavior reduces the amount of communication but affects the recovery accuracy since only few nodes are activated to track the object. On the contrary, in TORS, more sensor nodes are involved in the recovery process, which implies that more energy is dissipated and high accuracy is provided.
A prediction-based energy-efficient target tracking protocol (PET) was proposed in [
PET scheme suffers from a high missing rate when the node’s sensing range is small. By way of explanation, the smaller sensing range is, the more uncovered zones appear in the network and the higher probability of losing the object is.
Authors in [
In this approach, the frequency at which the prediction mechanism is invoked depends on the object’s movement. When the object moves at high speed, the prediction algorithm is called several times to estimate the target’s position. The problem with this scheme is that it uses a very complex prediction algorithm that consumes larger energy.
A Dynamic Lookahead Spanning Tree Algorithm (DLSTA) was proposed in [
The object tracking tree performance is affected when the movement behaviors of the object are different from the mobility profile already predefined. To tackle this issue, authors in [
MTA procedure incurs very high overhead to enhance the update cost of the object tracking tree. The situation gets worse when the size of the tree increases, which means that more adaptive messages will be transmitted to the sink along the object tracking tree.
Probability-Based Target Prediction and Sleep Scheduling (PPSS) protocol [
PPSS assumes that the object moves in a smooth curvilinear trajectory. However, this approach suffers from a main limitation which is the incapacity of handling abrupt changes in target direction. As a result, PPSS’s performances are degraded since it will make a long delay to compute the target’s position.
Authors in [
However, MCTA keeps all sensor nodes in the contour turned on without using a sleep scheduling mechanism to alternate between active and sleep modes. As a consequence, more sensor nodes are awakened which means more energy is consumed.
Since MCTA focuses on vehicle’s kinematics, it seems obvious that it will fail to track the other types of targets like tanks, unmanned aerial vehicles (UAV), and so on.
Authors in [
Chen et al. [
DSA2 is not efficient when it is about recapturing the target after it gets lost. The missed object is located by sensor nodes that happen to be in the sensing mode.
Data association (shown in Figure
Data association problem
Authors in [
This approach focused on tracking a group of targets that move collectively and have correlated motion, speed, and direction. It is exactly the case of a group of wild animals that migrates from one habitat to another. The problem with this scheme is that it does not handle scenarios where these sets of targets (animals) split, merge, or even overlap.
A Collaborative Boundary Detection and Tracking of Continuous Objects in WSNs was proposed in [
CODA [
When sensor nodes detect the presence of a target, they transmit their detection information to their immediate cluster head. Based on this information, the CH identifies the boundary sensors within its cluster. This process will cost additional energy, especially when the target overlays a large number of nodes.
A comparison between different target tracking approaches is drawn in Tables
Classification of target tracking approaches.
[ |
[ |
[ |
[ |
[ | |
---|---|---|---|---|---|
Network structure | |||||
Face | — | — | ✓ | — | — |
Tree | — | — | — | — | — |
Cluster | |||||
S: static | S&D | S&D | — | D | — |
D: dynamic | |||||
Prediction | — | — | Linear mobility | — | Two-dimensional Gaussian distribution |
Type of target | |||||
Continuous | ✓ | — | — | — | — |
Discrete | — | ✓ | ✓ | ✓ | ✓ |
Number of targets | |||||
Single | — | ✓ | ✓ | ✓ | ✓ |
Multiple | — | — | — | ✓ | — |
Binary sensor | — | — | — | — | — |
Recovery | — | — | TORS & CORS | — | ✓ |
Classification of target tracking approaches.
[ |
[ |
[ |
[ |
[ | |
---|---|---|---|---|---|
Network structure | |||||
Face | — | — | ✓ | — | — |
Tree | — | — | — | — | ✓ |
Cluster | |||||
S: static | S | — | — | — | D |
D: dynamic | |||||
Prediction | ✓ | Kinematics & theory of probability | ✓ | — | Linear, EKF, and PF |
Type of target | |||||
Continuous | — | — | — | — | — |
Discrete | ✓ | ✓ | ✓ | ✓ | ✓ |
Number of targets | |||||
Single | ✓ | ✓ | ✓ |
|
✓ |
Multiple | — | — | — | ✓ | — |
Type of sensors (binary) | — | — | — | ✓ | — |
Recovery | — | — | ✓ | — | ✓ |
Classification of target tracking approaches.
[ |
[ |
[ |
[ | |
---|---|---|---|---|
Network structure | ||||
Face | ✓ | — | — | — |
Tree | — | Tree combined with Vornoï Graph | — | — |
Cluster | ||||
S: static | — | — | — | S&D |
D: dynamic | ||||
Prediction | LSM least square | — | Kinematics | — |
Type of target | ||||
Continuous | — | — | — | — |
Discrete | ✓ | ✓ | ✓ | ✓ |
Number of targets | ||||
Single | ✓ | ✓ | — | ✓ |
Multiple | — | — | ✓ | — |
Type of sensors (binary) | — | — | — | — |
Recovery | ✓ | — | — | ✓ |
Classification of target tracking approaches.
[ |
[ |
[ |
[ |
[ | |
---|---|---|---|---|---|
Network structure | |||||
Face | — |
|
Uses |
— | — |
Tree | — |
|
— | — | |
Cluster | — | S | S |
|
|
Prediction | — | KF | Kinematics | ✓ |
|
Type of target | |||||
Continuous | — | — | — | — | ✓ |
Discrete | ✓ | ✓ | ✓ | ✓ | — |
Number of targets | |||||
Single | — | — | — | ✓ | — |
Multiple | ✓ | ✓ | ✓ | — | — |
Type of sensors (binary) | ✓ | — | — | ✓ | — |
Recovery | — | — | — | — | — |
Balancing the tradeoff between energy conservation and tracking accuracy is a big challenge. The more sensors are involved in tracking target, the more accurate the tracking is. However the use of a large number of sensor nodes in the tracking process causes high energy dissipation.
Most of the approaches introduced in this paper try to minimize the energy consumption while enhancing the tracking accuracy by using several techniques such as clustering, prediction of target’s movement, adopting specific network structures, and utilizing binary sensors.
There are several reasons that hamper the tracking and lead to target loss: Communication failures, presence of obstacles, and coverage holes. Abrupt change in target’s velocity and direction. Sensor nodes driven by batteries and having low-energy resources (when they remain in active mode for a long period, they deplete their energy faster causing the so-called “energy hole” problem [ Inaccuracy while estimating the target’s actual position. The long delay that a predicted algorithm may take to compute the future position of the target.
Unfortunately, HCTT, MTA, [
In terms of energy efficiency, PPSS, MCTA, and [
Tree-based approaches are a good choice for target tracking. However, they induce a high overhead cost. In tree architecture, lots of messages are exchanged in the network from leaf nodes toward the sink. This situation becomes more complicated, especially when trees are large. Furthermore, sensor nodes located near the root consume much energy in comparison with the other nodes. For instance, in the MTA scheme, many messages are exchanged between the sink (root node) and leaf nodes along the tree, causing a high overhead. MTA approach is more practical when the size of the network is small.
FOTP, POOT, and [
The VF approach achieves high energy efficiency since it uses binary sensor nodes in tracking. These so-called binary sensors generate one bit of information regarding the presence or absence of the target in their sensing range which reduce the amount of data transmitted in the network. However, this type of sensors cannot provide enough information about the chased target (velocity, direction, etc.).
DLSTA scheme consumes less energy and provides accurate tracking because it considers targets that move at high speed and make nonlinear and linear motions. In contrary to the other approaches, it employs three types of prediction techniques, linear, PF, and EKF, which make the DLSTA a strong approach. Moreover, since DLSTA have a smaller miss ratio, the recovery algorithm is executed rarely and thus more power is saved. By contrast, the recovery process in the adaptive head [
Much additional research effort will be needed to handle the problem of data association in multiple target tracking, to treat the problem of inaccurate measurements and false alarms that may occur in the network, to track mobile objects in presence of obstacles, since sensor nodes are prone to failures due to energy depletion, communication errors, and malicious attacks. More research efforts should be done to deal with fault tolerance issue in target tracking. None of the protocols listed in the tables below have considered it except the approach in [
Tables
Target tracking has gained considerable attention in recent years for its application in different fields such as military, civilian, and wildlife monitoring. The massive research in this field has inspired us to present this survey. In this paper, we investigated some of the target tracking algorithms currently used in wireless sensor networks. We compared and analyzed them from different angles. It is clear that all protocols presented in this paper share one common objective, which is ensuring a high target tracking accuracy while maintaining the energy.
The authors declare that there is no conflict of interests regarding the publication of this paper.