280 likes | 419 Views
A Sensor Placement Approach for the Monitoring of Indoor Scenes. P. David, V. Idasiak, F. Kratz Laboratoire Vision et Robotique, UPRES EA 2078 ENSI de Bourges - Université d'Orléans 10 boulevard Lahitolle, 18020 Bourges Cedex, France. SUMMARY. Background Problem Definition
E N D
A Sensor Placement Approach for the Monitoring of Indoor Scenes • P. David, V. Idasiak, F. Kratz • Laboratoire Vision et Robotique, UPRES EA 2078 • ENSI de Bourges - Université d'Orléans • 10 boulevard Lahitolle, 18020 Bourges Cedex, France
SUMMARY • Background • Problem Definition • Reused Works • Modelling Method • Genetic Algorithm • Case Study • Conclusion & Future Works
BACKGROUND • Development of a new human presence sensor • Creating a sensor simulator for the product being designed: • Testing design choices • Proving the performances of the system • Helping the deployment of the system • Finding guidelines for the implantation
PROBLEM DEFINITION • Sensor placement in indoor scene (housing, office) • Graduated importance for the monitored area • Avoiding to monitor parts of the scene • Selection of sensors (type, settings, capabilities) • Limited places for the sensors • Limited number of sensors The problem is to find the best sensors placement and selection to cover a scene with limited resources and heterogeneous goals.
REUSED WORKS 1/2 • Mainly inspired by works on video surveillance: • Similar goals (monitoring the activity of human) • Same kind of observed scenes • Basis brought by Erdem & Sclaroff [16]: • Method to convert a coverage problem as a linear programming problem • Finding solution for an entire monitoring of a room with a minimum number of cameras • Easy to reuse and enhance way of modelling
REUSED WORKS 2/2 • Weaknesses of Erdem & Sclaroff ’s solution: • Not considering aspects as price and energy consumption of the camera network • Considering only camera sensor, only one type of camera • Considering only the research of the total monitoring of the scene • Not considering priorities between different zones, or zones to avoid • Not considering balancing the efficiency and the cost of the sensor networks
Modelling Method EXPECTED DATA • Input data: • Geometry of the scene • Geometry of the furnishing • Usable sensors and possible location • Monitoring objectives • Output data: • Selection of sensors • Indicators (ex: coverage, cost)
Modelling Method TO A RESOLUTION PROCESS 1/2 • First Mathematical Formulation: • Minimize f(x) = cT x • With respect to A x ≥ b • ub x lb • x is a binary vector • ub and lb are respectively a vector of 0 and a vector of 1 • Easy to optimise with a Branch and Bound algorithm • Build matrix A and vectors b and x to represent the problem
Modelling Method TO A RESOLUTION PROCESS 2/2 • The vector x is the solution: the selected sensors • The vector b represents the constraints: the scene description • The matrix A allows to compute the effectiveness of a given sensor network • The function f is to be minimized: the price of the network (hardware price + energy cost) • x is a selection vector giving the indices of the selected sensors • A is the concatenation of all the vectors representing the information provided by each sensor, presented in the same shape as b
Modelling Method SCENE DESCRIPTION 1/2 • The basic vector b is constructed as follows: • The scene is sampled in a list of point numbered with growing indices. • The number of points gives the vector size. • All elements are set to 0. • If a point belongs to the zone that we want to monitor the corresponding element of b is set to 1.
Modelling Method SCENE DESCRIPTION 2/2 • For a more complex model including importance graduation we suggest to add a second vector giving the level of priority (Є ) for the detection.
Modelling Method SENSOR DESCRIPTION 1/3 • To characterize a sensor we need to know its efficient zone geometry and its implantation point. • This description can be completed by the tracked flow and the measure’s reliability. • The first step is to identify the geometry of the efficient zone of the computed sensor. • The second step is to compute this zone by ray tracing • Then the vector representing the sensor is computed • The efficient zone is the zone in which the sensor can provide information on the measured flow
Modelling Method SENSOR DESCRIPTION 2/3 0 0 .. . 0 0 1 1 . . 0 1 1
Modelling Method SENSOR DESCRIPTION 3/3 • The vectors of each sensor are concatenated in the A matrix so that A x is the vector representing a sensor network selected by x. • A x is therefore the list of points viewed by the network • In order to deal with more complex aspects as reliability, it is possible to compute a second vector indicating for each point the reliability of the measure given by a sensor. • The employed reliability law should be of any form describing a spatial distribution. (ex: Linear, exponential)
Modelling Method THE COST FUNCTION • The function is a simple vector multiplication. We created a vector which indicates the cost of the installation and exploitation of each sensor. • By multiplying this vector by the selection vector we obtain the cost of the network. This cost is then to be minimized to find the best solution
Modelling Method PROBLEM DESCRIPTION ≥
Modelling Method A FIRST SOLUTION Example of solution with a Branch and Bound Algorithm
Modelling Method FIRST OBSERVATIONS • + Easy to compute • + Easy to solve with a Branch and Bound algorithm • + Taking into account price constraint, placement constraint. • + Give the guidelines to create a more complex data structure • Many parameters are not optimised (financial and technical efficiency) • Complex constraints as priorities in the monitoring are not tackled • Use of genetic algorithm with a stronger data structure
Genetic Algorithm EXPECTED ADVANTAGES • The use of genetic algorithms is justified by the following points: • Optimise heterogeneous properties of the system (ex: cost, efficiency, redundancy) • Taking into account heterogeneous constraints (ex: zone to monitor, to avoid, cost limitation) • Those algorithms allow us to find solutions respecting more complex wishes than finding a system monitoring a complete scene whatever its price. • This gives results more accurate for real exploitation
Genetic Algorithm ENHANCE DATA STRUCTURE • Computing new matrices similar to the A and Scene matrices. • Inserting priority of observation in a second Matrix representing the scene. • Adding a second column for each sensor in the sensor Matrix A, indicating the viability of the measure given in the designated point. • Associating complex cost involving energy consumption, maintenance cost and hardware price.
Genetic Algorithm BASIC FITNESS FUNCTION val 0 for i=0..NbPoint if Scene ( i ) > 0 if (A x)( i ) ≥ Scene ( i ) val val + 100 / sum(Scene ) if (A x)( i ) < 0 val val – 100 if sum( x ) > MaxSens val val – 100 val val + Cov ( NbSens – sum( x ) )
Case Study THE MONITORED SCENE
Case Study POSSIBLE SENSORS LOCALISATION
Case Study RESULTS
CONCLUSION & FUTURE WORKS • First development of our placement tool. • Use genetic algorithm to provide a more realistic optimisation. • Possibility to model more complex implantation policies. • Taking into account the third dimension of the scene (Full 3D or using 3 parallel planes at significant heights). • Adding zone of perturbation for several types of sensors (US, PIR, …).
CONCLUSION & FUTURE WORKS • Implementing a GUI allowing to load plans of the scene and to choose the shape and parameters of the fitness function. • Integrating the final tool in a complete simulation tool. • Integrating, in the results, the enhancement brought by data fusion methods. • Integrating the resolution of the measure to know what kind of treatment could be performed on every point of the scene. • Repeating experiences in various environments to extract general guidelines of placement.