A key motivating application for group-based duty cycling is to understand the behaviour of migratory birds [24]. In particular, flying foxes are major cause for disease spread in both Australia and in Southeast Asia. These animals are highly mobile, traveling tens of kilometres in a single night and covering distances in excess of 1,000 km with the migratory season. In Australia, flying foxes are responsible for the spread of the Hendra virus, which is lethal to horses and harmful to humans. The processes and dynamics that underpin disease propagation within a group of flying foxes are not well understood, which is the main driver to monitor their movements and understand their behavior. Fortunately, these animals congregate at roosting camps and groups of tens of thousands of individuals and may remain there for days or weeks before moving to a new location. Using group based duty cycling therefore has huge potential for developing long-term tracking solutions for these animals. A related application is monitoring of livestock that also tends to form into long-term stable herds. The benefits of collaborative localisation for cattle have been demonstrated in [5] through reliance on radio proximity data to reduce GPS duty cycles. In that earlier work, nodes did not have any inertial sensors. The animal scientists that wish to study the cattle behaviour are interested in high-quality inertial data from the animals while they are moving.

Both of the above applications require the use of inertial modules with tilt compensation and heading computation, which have a much higher energy consumption than basic inertial modules. The need to use higher energy inertial modules motivates duty cycling of these modules. In order to avoid missing motion events, multiple nodes that are clustered together in a group can cooperate on detecting motion in the group and simultaneously save energy.

We consider a set of nodes N = {N_i}, where N_i denotes an individual node. Our goal is to track location L_i of each node over a set of discrete timestamps with a maximum user-set uncertainty. In order to ground the discussion, we consider accelerometers as our motion trigger modules, bearing in mind that the model applies equally to other motion sensors, such as gyroscopes and magnetometers.

Each node is equipped with a radio that has a perfectly circular range of R meters. The nodes are further equipped with an accelerometer for motion detection, and a GPS module that estimates the node’s location L_i with an average uncertainty U_gps meters. An accelerometer takes t_A time to estimate the node’s current speed, and the GPS module requires t_L seconds to achieve lock and return a location estimate. Because accelerometer readings are noisy and cannot reliably estimate fine-grained speed values, we use the accelerometer in our model to distinguish between two states: a stationary state and a moving state. The accelerometer can determine whether movement occurs through an internal mechanism that compares the measured speed with a user set speed threshold

We consider an outdoor cattle tracking application using smart collars that contain wireless sensor nodes, GPS, and inertial modules with accelerometer, magnetometer, and hardware tilt compensation and heading calculation. The wireless sensor node is a Fleck [26] comprising Atmel-1281 MCU and Atmel RF212 radio. Four D-Cell batteries in series provide power to all the collar components via several switch-mode regulators. The weight of these batteries yields a collar weight that approaches the upper limit set by an animal ethics committee, so larger batteries are not an option.

The target node lifetime in this application is 3 months, which is the interval at which the animals are brought in for health checks, treatment and sorting. Previous work had shown a significant increase in node lifetime with GPS duty cycling coupled with radio ranging [5]. As our current work on collaborative localisation with accelerometer duty cycling builds on that earlier work, we briefly revisit it for presentation clarity.

Figure 1 shows how GPS duty cycling works. At time k we obtain a GPS measurement x_k and then turn off the GPS. At the next measurement x_k+1, the mobile node may be either within or outside the AAU, indicated by the larger circle. Whenever x_k+1 falls outside the larger circle, we denote this as an error, since the system would have failed to deliver a position estimate within the AAU.

This approach requires a means to estimate the uncertainty bound (shown as a dashed circle) at any given time. Since we do not know how fast or in what direction are the nodes moving, the simplest model is to assume a certain maximum speed and grow the estimated uncertainty linearly with time. When it approaches the AAU, taking into account the GPS’s lock time, we turn on the GPS. If the node’s speed exceeds our assumption, it could end up at the point marked in red when the GPS obtains a fix. This paper uses the dynamic speed model, which updates the assumed speed of a node every time it obtains a GPS lock [5].

GPS lock can take several seconds, during which the module obtains an initial lock with a relatively high U_gps, and then exponentially decreases its uncertainty as it receives more signals from visible satellites [5]. Keeping the module on for longer yields higher position accuracy, but it incurs heavy energy cost. With the GPS duty cycling algorithm, a target GPS uncertainty can be set so that the total off time of the module (considering both the time spent obtaining lock and the time waiting for the algorithm uncertainty to grow) is minimised. Through offline calculations on the model in [5], we determine this target GPS uncertainty to be 13 m for an AAU of 50 m. We use these settings for the rest of this paper for more aggressive GPS duty cycling.

Alongside GPS duty cycling, nodes use radio beacons to provide local position references. The beacons also include the sending node’s uncertainty. Whenever a node receives a beacon from a neighbour with a lower uncertainty, it adopts the position estimate of that neighbour and reduces its uncertainty according to its estimated distance from the neighbour. In order to receive radio messages, nodes use a low-power listening mechanism [6]. Low power listening keep nodes’ radios in low-power sleep states for most of the time. The radio periodically wakes up for a short channel check assessment (CCA) time to check if there are any upcoming packet transmissions for this node. If so, the radio remains on to receive those packets. Otherwise, it goes back into sleep state. We use the same low-power listening approach for all radio beacons in this paper.

While the earlier work focused on bounding the location uncertainty by growing it linearly over time according to an assumed speed [5], accelerometers enable nodes to keep their uncertainty constant over long periods and thus reduce the duty cycle of the node’s MCU and GPS units. However, simply including an always-on accelerometer within the node is not energy-efficient, particularly for modules with heading and tilt compensation.

Table 1 shows the power consumption of the various node components in both active and sleep modes. Our earlier work had shown that GPS Duty cycling with radio proximity can achieve a combined power consumption for the GPS, radio, and MCU around 22 mW. In comparison, the inertial module would consume 18 mW if it remains in active mode all of the time, representing an 82% increase in overall power consumption for the node. While we expect the inclusion of the inertial module to reduce the duty cycles and energy consumption of the other node modules, any energy savings are likely to be outweighed by the increased energy consumption for powering the inertial module. Therein lies the need for aggressive duty cycling of the inertial module. The following sections explore the extent to which individual accelerometer duty cycling can improve the previous GPS/radio ranging strategy.

This strategy enables nodes to set their accelerometer duty cycle independently of any neighbours. The individual duty cycling approach follows a state-based approach with two states: motion state St_M and stationary state St_S. The node uses the accelerometer to detect motion and defines St_M state for speeds above threshold and St_S for speeds lower than the threshold. The individual duty cycling algorithm is shown in Figure 2, while Figure 3 shows a typical duty cycling scenario.

We augment the python simulator in [5] with the accelerometer duty cycling algorithm and energy model to evaluate their performance with our empirical cattle GPS data trace. Figure 6 shows a trace of the uncertainty growth for 3 nodes from our empirical cattle dataset. The trace uses solid lines to indicate the algorithm uncertainty and dashed lines to indicate the true error for each node. At the start of the trace, all three nodes are growing their uncertainty linearly according to , after having logged contact with a common neighbour 104, whose trace is not shown here. All of the narrow peaks in the trace indicate a node detecting motion after sampling its accelerometer, and adding a step increase to its uncertainty. The reason that the uncertainty drops back again is that the node mobility is transient and the nodes receive periodic contact beacons from neighbours that have a lower uncertainty shortly thereafter, which enables the nodes to lower their own uncertainty. It is specifically for this reason that the individual duty cycling algorithm has to use periodic radio beaconing (rather than event-based beaconing), in order to avoid premature GPS locks due to short-lived motion events.

Interesting situations arise at times 400 and 1,500. At 400 s, nodes 299 and 337 both detect motion independently, but in this case, they cannot lower their uncertainty after they leave the motion state because they do not receive any contact beacons with sufficiently low uncertainty. This situation persists until a neighbour acquires GPS lock, which enables all three nodes in the trace to lower their uncertainty by logging contact with that neighbour. At time 1,500, node 337 first detects brief motion, which causes a spike in its uncertainty. Once its accelerometer detects that it has stopped moving, it returns to a slow linear increase in uncertainty thanks to contact logs with node 299. A similar pattern occurs at node 103 about 100 s later. After that, node 299 detects that it is moving, and adds a step increase to its uncertainty. Note that the actual position error remains within the uncertainty at all times.

We focus on two performance parameters: energy consumption and localization error of the duty-cycled algorithm. Specifically, we would like to minimise the energy consumption of each node while keeping the localization error within a maximum uncertainty AAU and the error rate σ within an error tolerance σ^* of 5%, because the animal scientists who are interested in the tracking data can accept a 95% success rate for the algorithm. If the duty cycle is low, we save more energy but increase the latency of motion detection, which can compromise the AAU uncertainty bound. All references to AAU error refer to the percentage of instances that the tracking algorithm exceeds the user-set maximum error.

In this section, we have introduced and evaluated an algorithm for duty cycling the accelerometer module on a mobile node based on local information only. We have also explored the design space for this algorithm and established the suitable parameter settings for our target scenario of energy efficient localisation within an error tolerance of 5%. The next section will use this configuration to propose group based duty cycling of accelerometers.

The utility of collaborative localisation depends highly on the clustering dynamics of the monitored individuals. Individuals whose movement is tightly coupled will benefit more from collaborative localisation than individuals that move independently of others. This paper defines clusters according to spatial proximity: nodes that are within a certain distance R of each other for at least T seconds are part of the same cluster. Naturally, clusters vary over time depending on the movement of the constituent individuals.

We define a boolean instantaneous proximity function:

Instantaneous proximity estimates can be highly noisy due to signal variations and for nodes at the boundary of the proximity radius. Localisation estimates, particularly those based on GPS, can be noisy and unreliable for single samples, and they tend to be more reliable when taken over several samples. These variations emphasise the need for capturing proximity over a time window rather than a single sample. Extending γ for a time window of n discrete time samples T = T₁…T_n yields:

This function returns a value between 0 and 1 that represents the proportion of samples in window T during which nodes N_i and N_j are within the proximity radius R. A proximity of 1 over a time window means that the two nodes have been within a radius R for the whole duration of the time window. Similarly, a γ_T of 0 means that the nodes are beyond the proximity radius for the whole duration of the time window.

The selections of R, T and the minimum threshold γ_T that governs cluster membership clearly have a direct impact on proximity. Larger proximity radii deliver large cluster memberships, while larger time windows and threshold γ_T provide more stable clusters. Both of these effects are desirable for collaborative localisation. However, both of these values involve inherent trade-offs. Using excessively high proximity radii for collaborative localisation means that nodes have to rely on cluster members that are farther away, which reduces the localisation accuracy. Similarly, time windows that are too long or γ_T thresholds that are too high can lead to smaller clusters because of their proximity requirement for a longer duration, which reduces the benefits of collaborative localisation.

We now explore these effects on our cattle dataset. In order to clearly separate transient and strong proximity, we are seeking proximity distributions that are well segmented. To that end, we run 16 analysis scenarios evaluating a range of values for the proximity radius R = 5, 10, 20, 30 m and for the time window length T = 5, 10, 30, 60 min. Each scenario runs through the entire datasets and determines the pairwise node proximity values for the given R and T settings.

Figure 14 shows the results of each of the radius threshold and time window value. Each row corresponds to one setting of the proximity radius R and each column corresponds to one setting of the time window. Each plot in the figure shows the histogram of γ_T values for the corresponding values of R and T.

As expected, a large proximity radius (20–30 m) and a short time window (5–10 min) provide the best segmentation. These settings also provide a high proportion of instances with very high proximity. At first glance, this is highly desirable. The following sections explore the effect of large proximity radii on collaborative localisation.

The main concept of collaborative duty cycling is to share the burden of motion detection among a group of nodes whose movement appears closely interdependent. To detect a motion event while the group is in a stationary state, at least one of the nodes in the group must sample its accelerometer during the event. Because motion events persist for at least T_M, at least one node in the group must sample its accelerometer every seconds. With G nodes in the group, each node can power off its accelerometer for seconds. Note that we only rely on collaborative duty cycling in the stationary state. When nodes are in the motion state, they revert to individual duty cycling. Figure 15 shows a typical motion detection time line with our group based duty cycling algorithm.

Figure 16 illustrates how collaborative duty cycling of accelerometers works alongside GPS duty cycling. All nodes send periodic radio beacons that contain their uncertainty and current number of neighbours, with period T_R, to maintain neighbourhood state. Each node maintains a neighbourhood table where it stores the proportion of neighbour beacons received in the recent time window T. Periodic beacons also serve the purpose of limiting neighbour uncertainty, as in [5]. Nodes that receive a beacon from a neighbour with a lower uncertainty can reduce their own uncertainty by using that neighbour as a reference.

Whenever a node obtains position lock from GPS, it samples its accelerometer for seconds. If it detects no motion during that period, the node transitions into stationary state. At this point, the node determines its true neighbours, which are the nodes in its neighbourhood table for which the node has received at least δ beacons in the past T seconds. The δ parameter is a tunable proximity threshold that defines the degree of level of proximity that provides a suitable degree of confidence in neighbour movement correlation.

The node also considers the number of neighbours of all of its true neighbours (recall that each node embeds its number of true neighbours in periodic beacons) and selects G = min(G(N)), where G(N) is the number of true neighbours of neighbor N. This step is required to ensure consistency in accelerometer duty cycling among nodes in the same group, when some of the nodes belong to overlapping groups of nodes. The node can then set its own duty cycle locally as . The reasoning is that each of the node’s true neighbours will run the same local computation and set at least the same duty cycle (depending on their neighbours’ number of neighbours). Determining true neighbours and setting the collaborative accelerometer duty cycle repeats at every time step while the node is in stationary state to ensure adaptation to neighbourhood changes.

Whenever any node samples its accelerometer and detects movement at a speed above , the node leaves the stationary state and broadcasts a warning beacon to all of its neighbours. All neighbours that receive the warning beacon sample their accelerometer immediately to determine if they too have started moving. If so, the neighbours also enter the motion state, and otherwise, they stay in the stationary state.

Nodes that enter the motion state increase their uncertainty by , as in the individual duty cycling case. As long as their uncertainty has not reached the critical limit of AAU − t_L, they duty cycle their accelerometer less aggressively with an off time of and update their uncertainty every time instance by . If the uncertainty reaches the critical limit, the node powers on its GPS for a new position estimate. If the node powers on its accelerometer before the uncertainty reaches the critical limit and detects no motion for at least seconds, it transitions back into stationary state.

Figure 17 shows a trace of the uncertainty growth for 3 nodes from our empirical cattle dataset. The trace uses solid lines to indicate the algorithm uncertainty and dashed lines to indicate the true error for each node. At the start of the trace, all three nodes are growing their uncertainty linearly according to , after having logged contact with a common neighbour 104, whose trace is not shown here. At around t = 200, node 103 detects motion through its accelerometer, which results in a step increase in its uncertainty. However, in the following time instance, node 103 logs contact again with node 104 and reduces its uncertainty again. This happens again in the next cycle with a smaller peak, as node 103 has now switched to a shorter accelerometer duty cycle in motion state. After that, node 103 detects that it has stopped moving, and reverts to slow uncertainty increase. Nodes 337 and node 103 are in the same situation at t = 1,200, 1,250 respectively. At around t = 1,225, node 299 has a sharp increase in uncertainty before node 337 obtains GPS lock and lowers its uncertainty. This is a consequence of using the dynamic speed model with step increases in uncertainty upon motion detection. The last time node 299 got GPS lock, it had a speed of 0.8 m/s. It keeps this speed estimate until its next GPS lock. When it detects motion, it adds to its uncertainty, which causes this sharp rise. When node 337 obtains lock a few seconds later, this situation is rectified. Another node 394 (not shown here) is also nearby and obtains lock at nearly the same time. At around 1,450 s into the simulation, node 337 detects motion on its accelerometer and sends a warning beacon to its neighbours. Node 103 detects that it is also moving after sampling its accelerometer, while node 299 detects that it is not. Both nodes 103 and 337 add a step increase to their uncertainty after detecting motion. They shortly log contact with node 394 to reduce their uncertainty again. Note that the actual position error remains within the uncertainty at all times.

The accuracy of partitioning nodes into groups that travel together critically influences both the power consumption and error rate metrics of the localisation algorithm. Nodes use their energy unnecessarily if triggered by a motion beacon when they are not moving. On the other hand, the localisation error rate grows when the nodes fail to update uncertainty bounds as a result of not receiving a motion beacon in the motion state.

We compare the performance of individual duty cycling, group-based duty cycling with 2 baseline scenarios: (1) an always-on accelerometer that triggers a duty cycled GPS to turn on when there is motion, where radio contact logging is used for local position sharing; and (2) the previous work on GPS duty cycling and radio proximity, with no accelerometer. We run two separate sets of simulations for these approaches with different travel times, in order to illustrate how performance can be tuned to favour energy or error rate. The algorithm with no motion accelerometer uses event-based beaconing, which had been established as the most effective approach in [5], while all other algorithms use periodic beaconing, for reasons discussed in Section 4 and Section 5.

Figure 21 compares the results. Both flavours of our accelerometer duty cycling strategies reduce the power consumption by around 50% over the GPS duty cycling with radio ranging approach and 72% over the always-on accelerometer approach. This comes at the cost of an increased error rate that remains within our application’s 5% error tolerance. The higher error rate with accelerometer duty cycling is intuitive, as nodes slow down their uncertainty growth in stationary state until they detect movement. The bars for a travel time of 15 s show the performance of accelerometer duty cycling for applications with tighter accuracy constraints. This shorter only causes minor increases in power consumption (about 1 mW for individual duty cycling and less than 0.5 mW for collaborative), while reducing their error rates to 2% and 3.5% respectively. Group duty cycling saves about 10% in power consumption over individual duty cycling, as nodes can set longer accelerometer off time. This comes at an increase of 1.25% to the error rate, since the longer accelerometer off times increases the likelihood of missed motion events.

Notably, the algorithm with no accelerometer consumes significantly higher energy for powering the GPS module and the MCU, while consuming less than half the energy for radio communication. The higher GPS power consumption is clearly because of the more frequent GPS fixes in the absence of accelerometer triggering. The MCU power consumption is also higher because the MCU has to remain active while the GPS is powered on (see overlap parameter in energy model). The radio power consumption is lower for this algorithm as it uses event-based beaconing, while both individual and group-based duty cycling use periodic beaconing. The results clearly show that the added cost of periodic beaconing is well worth the savings in GPS and MCU power consumption, which result from a significant decrease in the frequency of GPS fixes.

While keeping the accelerometer always on to trigger GPS position locks does reduce GPS power consumption thanks to a reduced frequency of locks, it is the most energy-expensive option, as the accelerometer module consumes 18 mW to power up its tilt compensation and heading calculations in hardware. This algorithm also results in the best location accuracy, slightly outperforming the no-accelerometer algorithm due to more adaptive GPS locks.

This section examines the performance dependence of our duty cycling strategies on the degree of mobility.

We first evaluate the effectiveness of our duty cycling strategies for highly mobile scenarios. We select the 3-hour time window from the data trace with the highest number of motion events, and we run each of the duty cycling strategies for all the nodes during that time window only. The results are shown in Figure 22. Both individual and group-based accelerometer duty cycling strategies reduce the power consumption relative to GPS duty cycling with radio proximity (between 21% and 35% respectively) and relative to the always-on accelerometer (between 60% and 62% respectively), at the cost of increased error rates as they use accelerometer estimates to reduce reliance on GPS. Both approaches deliver comparable error rates with GPS duty cycling while delivering this energy benefit. For the group-based strategy, a 30 s caused the error rates to go above the error tolerance, so we show here the results for of 15 s and 25 s for that strategy. The individual accelerometer duty cycling approach outperforms the always-on approach, indicating that duty cycling the accelerometer can lead to fewer false positives in detecting transient motion.

Next, we explore the case of rare movement events. We select the 3-hour time window from the data trace with the least number of motion events, and we run each of the duty cycling strategies for the all the nodes during that time window only. The results are shown in Figure 23. Both individual and group-based accelerometer duty cycling strategies significantly outperform GPS duty cycling with radio proximity (50%–65%) and the always-on accelerometer (65%–78%) in terms of power consumption, with a modest increase in error rates that remains well within the application-specific limit of 5%. The group-based approach delivers a further 20% improvement in power consumption over individual duty cycling, as the nodes that share the accelerometer sampling burden get to spend a longer time in off mode. Interestingly, a of 30 s yields the best results for low mobility, as it is less likely to miss motion events during the longer accelerometer sleep time. GPS duty cycling with radio ranging delivers the highest accuracy, outperforming the always-on accelerometer. All versions of accelerometer duty cycling exhibit slight reductions in accuracy, while staying well within the 5% error tolerance.

The results in this section have shown that while both flavours of accelerometer duty cycling improve the power consumption over previous work, sharing the burden of duty cycling among members of a group is beneficial primarily in situations where the expected likelihood of motion is relatively low. When that is not the case, performance of the two approaches is comparable with a slight improvement in error rate for the individual approach.

The results have shown that individual accelerometer duty cycling performs better during high activity time windows, and that collaborative localisation performs better during low activity time windows. This effect stems from the highly deterministic and synchronous nature of individual duty cycling. For the same reason that synchronous MAC protocols perform better with high traffic scenarios, individual duty cycling is well-suited for situations where the likelihood of a motion event is high. The reason that it outperforms collaborative duty cycling in this situation is that the latter distributes the burden of activity check across multiple neighbours, some of which may leave the group during critical times.

Through similar reasoning, collaborative duty cycling performs well under low activity scenarios as the dynamic nature of moving nodes adds a randomisation effect to the sample schedules. This randomisation is beneficial for detection of motion events in the group, whose distribution in non-deterministic.

An interesting extension of our work is to enable nodes to adapt their accelerometer duty cycling strategy based on their observed motion levels. When the frequency of motion events exceeds a certain threshold, nodes switch to individual duty cycling. When activity slows down, they can revert to collaborative duty cycling. The results in Figure 22 and Figure 23 suggest that additional savings can be achieved by actively enforcing this strategy.

The group structure across a dynamic topology of mobile nodes will almost certainly lead to asymmetric views of groups among nodes. For instance, a node A may have ten neighbours, nine of which are tightly connected and have a consistent view of their group. These neighbours can simply set an accelerometer duty cycle at one-tenth of their individual duty cycle. Node A’s tenth neighbour, node B, only sees node A as its neighbour, because it is further away from the other nodes in the group. As a result, node B sets its duty cycle at 50% of its individual duty cycle. Node A will also set its duty cycle to 50% of the individual duty cycle, thanks to the min(G(N)) operation through which it can accommodate its least connected neighbour. This approach also holds true in the case of multiple group overlaps.

Another issue is whether group members are implicitly synchronised through warning beacons, and whether this has an impact on the performance of our collaborative duty cycling strategy. A few features in our algorithms help avoid implicit synchronisation. Firstly, nodes notify neighbours when getting GPS lock, which avoids race conditions for this operation. Second, nodes in overlapping groups will check their accelerometers at different intervals, which contributes to desynchronisation. Additionally, the use of the dynamic speed model leads nodes to grow their uncertainties at different rates in motion states, helping to avoid implicit synchronisation. Finally, the specific individual movement or non-movement patterns also help desynchronisation of uncertainty growth. A node in a group that detects movement at one instance will trigger a warning beacon to neighbours. A subset of neighbours will switch to motion state if they are moving as well. This will cause desynchronisation among nodes in the moving subset and the rest of the group. At another point during the process, another subset of the group may similarly detect motion and shift their uncertainty growth away from the rest of the group.

This paper has considered a simple mobility model for the animals consisting of a stationary and moving state. While this model can easily be generalised to other moving objects, more fine-grained modelling of the motion states can potentially yield improved accuracy and energy efficiency. In the case of cows, previous studies suggest that motion can be modelled using three [27] or four states [28]. The potential benefits of more fine-grained motion modelling should be weighed against the ability of motion detection modules to accurately distinguish between the states. This remains an open issue for future work.

The group based motion detection algorithms in this paper have used empirical data traces to explore the design space and establish the utility of this approach, particularly when the mobile entities move rarely. An interesting direction for future work would be to empirically test this approach in order to better quantify its benefits for real-time tracking. Empirical validation would require the use of cameras with targeted image processing algorithms for cattle tracking or through human observers for the duration of the experiment, which would limit the duration or spatial scale of the experiment.