Evolutionary Approaches for Resilient Surveillance Management Ruidan Li and Errin W. Fulp WAKE FOREST U N I V E R S I T Y Department of Computer Science BioSTAR Workshop, 2017
Surveillance Systems Growing interest in surveillance systems Decreased camera costs make these system affordable Applications include monitoring security and improving safety Management has become more complex Given a set of cameras how should they deployed? What happens if the environment changes? Team Wake NSG@WFU BioSTAR 17 1
Art Gallery Problem Determining camera locations and orientations is a well known problem Art Gallery Problem (AGP) originally introduced in 1973 Locate guards(cameras) in an art gallery(polygonal space) such that every interior wall was observed by at least one guard. Heuristics exist for AGP, but assume everything is static Team Wake NSG@WFU BioSTAR 17 2
Surveillance Problem Let s consider a variation of the AGP B α β A C α β C A Assume cameras are already located within the floor space Each camera has fixed view angle α and adjustable rotation angle β Can also turn a camera on or off Objective: Determine the smallest set of cameras and their orientation that maximizes wall coverage Team Wake NSG@WFU BioSTAR 17 3
Configuration Search Consider all the possible configurations for a floor plan and camera set Can apply a search algorithm, but perhaps not a traditional approach Team Wake NSG@WFU BioSTAR 17 4
Surveillance Resilience Cameras and/or obstacles may be introduced and/or removed 1 1 1 2 event add 1 2 event drop 2 2 0 0 0 Management should automatically adjust to changes Team Wake NSG@WFU BioSTAR 17 5
Evolving Surveillance Management gen 0 gen 1 gen 2 gen 3 gen 4 gen 5 gen 6 gen 7 gen 8... fitness 6.0 fitness 12.0 fitness 13.0 fitness 18.0 fitness 21.0 fitness 24.0 fitness 29.0 fitness 33.0 fitness 35.0 Evolutionary Algorithms (EAs) are used as search heuristics Better solutions are created from good solutions EAs have the ability to constantly search If the problem changes, then the EA can adapt the solution Team Wake NSG@WFU BioSTAR 17 6
Evolutionary Algorithm Components Chromosome a model for a solution to the problem Fitness value for ranking, we seek better solutions Processes combining solutions to create new one Pool set of chromosomes (also called a population) Team Wake NSG@WFU BioSTAR 17 7
Chromosomes and Fitness A chromosome is a solution to the problem Camera i has two settings: p i = {1,0} and β i = (0,360) 1 95 2 80 s = {{1,95} 0,{0,0} 1,{1,80} 2 } 0 Therefore a surveillance configuration s is a chromosome Fitness consists of two objectives Maximize the coverage (percentage of the wall space observed) Maximize the number of inactive (unused) cameras Use Pareto ranking to determine the best configuration Team Wake NSG@WFU BioSTAR 17 8
Evolutionary Algorithm Processes + pool = M selection crossover mutation Selection identifies parents for new chromosomes (configurations) Crossover combines selected chromosome to create new chromosomes Randomly change traits (camera settings) Team Wake NSG@WFU BioSTAR 17 9
System Operation current pool (p) contains the new generation no, configure another camera start select and crossover from current pool (p) mutate s full? no yes, new chromosome generated p full? add chromosome to new pool (p ) evaluate chromosome (s) copy new pool to current pool yes, new pool (p ) contains n chromosomes... generation 1 generation 2 generation 3 generation 4 generation 5 Continually iterates and updates configurations based on environment. Team Wake NSG@WFU BioSTAR 17 10
EA + Beam = Hybrid Beam search is another search heuristic that maintains a population Explores only the best x% solutions in the current generation Best x% mutated to create next generation Beam can focus on a specific area within the search space Combining EA and Beam may provide better search breadth and depth Combining can be easily done across generations E E E E E B B E E E E E B B E E E... 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 generation Provides refinement and exploration (if the environment changes) Team Wake NSG@WFU BioSTAR 17 11
Experimental Results Considered 40-sided orthogonal polygonal floor plans 10 different 40-sided orthogonal polygons generated 80 cameras were randomly located within the floor plan Interested in evaluating the resiliency of surveillance management Considered random camera add/removal and obstacle add/removal EA, beam search, and hybrid management approaches compared Seek high coverage and high inactive camera percentages Team Wake NSG@WFU BioSTAR 17 12
Adding and Dropping Cameras 4 Drop (D) events followed by 4 Add (A) events Average Coverage percentage 100 90 80 70 60 50 D D D D 300 320 340 360 380 Generation A A A EA Beam Hybrid A Average Inactive camera percentage 90 85 80 75 70 D D D D 300 320 340 360 380 Generation A A A EA Beam Hybrid A Evolutionary-based adapted better than beam search Hybrid performs slightly better than EA Team Wake NSG@WFU BioSTAR 17 13
Obstacles Four 4-sided obstacles introduced at generation 300 Average coverage percentage 100 90 80 EA Beam Hybrid 70 0 100 200 300 400 500 600 Generation Average inactive camera percentage 80 60 40 20 0 EA Beam Hybrid 0 100 200 300 400 500 600 Generation Evolutionary-based adapted better than beam search Coverage is similar; however, evolutionary-based used fewer cameras Team Wake NSG@WFU BioSTAR 17 14
Conclusions and Future Work Camera-based surveillance systems are increasing in popularity Providing resilient management is difficult (Art Gallery Problem) Most heuristics do not consider dynamic conditions Evolutionary-based management algorithms have several advantages Able to dynamically adjust to changing system conditions Several areas for future work More experiments are needed to better understand hybrid operation Experiment with complex events (camera changes and obstacles) Prefer the technique to be more distributed Team Wake NSG@WFU BioSTAR 17 15
Turn back, you ve gone too far Team Wake NSG@WFU BioSTAR 17 16
Multi-Objective and Pareto Optimal Weighted sum can be used to scalarize multi-objective values All solutions that yield an good average are considered equal Surveillance Configuration Performance Inactive cameras percentage 50 40 30 20 10 65 70 75 80 85 Coverage percentage first second third A configuration s i said to dominate the another solution s j : if the solution s i is no worse than s j in both objectives and the solution s i is strictly better than s j for at least one objective. Team Wake NSG@WFU BioSTAR 17 17