Although mobile robots have on-board sensors to perform navigation, their efficiency in completing paths can be enhanced by planning to avoid human interaction. Infrastructure cameras can capture human activity continuously for the purpose of compiling activity analytics to choose efficient times and routes. We describe a cascade temporal filtering method to efficiently extract short- and long-term activity in two time dimensions, isochronal and chronological, for use in global path planning and local navigation respectively. The temporal filter has application either independently, or, if object recognition is also required, it can be used as a pre-filter to perform activity-gating of the more computationally expensive neural network processing. For a testbed 32-camera network, we show how this hybrid approach can achieve over 8 times improvement in frames per second throughput and 6.5 times reduction of system power use. We also show how the cost map of static objects in the ROS robot software development framework is augmented with dynamic regions determined from the temporal filter.
The low end of the filter is TS1 and the high end TS2. The time constant TS1 separates the signal from in-place activity and TS2 removes high-frequency, infrequent noise. Short-term, moving activity is found by subtracting MS1 from ML1 to rid the in-place activity (effectively a high-pass filter) and then applying a low-pass filter to rid infrequent noise. This combination results in band-pass filtering, (ML1(x, y, t) MS1(x, y, t)) FS2 MS2(x, y, t) (7) Note that the two filters at the high frequency end of Fig. 2 are not redundant
id: 0b2b595e065a9527a77e38cc4bc627b3 - page: 5
Finally, event detection is performed on short-term, in-place and moving activity to act as a gate on more computationally expensive processing such as object detection, Event(MS1,MS2) = 1, do object detection = 0, do nothing. (8) 3.5 Off-line Global Path Planning Most commonly, a path is planned that avoids human activity in time and space. It is less common to choose path segments of high activity, but we do this in the following way. In Fig. 3, the long-term storage contains motion statistics for each minute of the day Mp(f, t ), where subscript p indicates this is from a camera viewing path segment p. We time-collapse and binarize this as follows, Mp(f) = 1, if f(t ) = 0 for any t , 0 < t < 1440 = 0, otherwise.
id: d863bd2747ffa995876bd62fc14e996d - page: 6
Fig. 3. Different types of activity information used for different path planning tasks. The result M p can be thought of as a location attribute learned from people activity. If M p is 0, then no people travel through this location for whatever reason, so its reasonable for a robot to avoid it as well. Off-line path planning through any location at t is now a function of two activity-related values, Mp(t ) and M p . The activity-related cost for a potential path through locations pi for off-line planning can be written, Cost1({pi}, t ) =X i Cost(Mpi (t )), if all M pi = 1 =, if any M pi = 0. (10) 3.6 Real-time Global Path Planning Since real-time global path planning is performed just before the robot begins a path, there is current short-term activity information available as well as longterm information as shown in Table 1. It may make sense to weight the shortterm information higher than the long-term information {w1, w2}, although we do not explore that further here. For simplic
id: 6d550ba36f9fe47e271fd42b53d5e8cb - page: 6
The activity-related cost for a path through segments {pi} from off-line planning is, Cost2({pi}, t)) =w1Cost1({pi}, t = t) + w2 X N i=1 Cost(Mpi (t)). (11) The top line of equation 11 is the long-term activity cost at isochronal time t = t, which for real-time planning is the current time of path planning. The cost in the bottom line includes both in-place and static short-term activity. 9 Short-term moving activity in Table 1, which is captured from the on-board robot sensors for local planning, is outside the focus of this paper, but if captured it would be added to equation 11. 4 Experiments and Results 4.1 Scope of Experiments The focus of this paper is on efficient design of an activity filter and application to human-robot interaction. It is important to state what is outside the scope of this paper. Experiments showing the effectiveness of activit
id: f634e34227a69f84f0386a4e4ee5b8b1 - page: 6