FACIAL SOFT BIOMETRICS - Library of Ph.D. Theses | EURASIP
FACIAL SOFT BIOMETRICS - Library of Ph.D. Theses | EURASIP FACIAL SOFT BIOMETRICS - Library of Ph.D. Theses | EURASIP
44 4. SEARCH PRUNING IN VIDEO SURVEILLANCE SYSTEMSfurther elimination of categories, which limits large databases of subjects to a fraction of the initialdatabase, see Figure 4.1. In the context of this chapter the elimination or filtering of the employedcategories is based on the soft biometric characteristics of the subjects. The pruned database canbe subsequently processed by humans or by a biometric such as face recognition.The approach of pruning the search using SBSs, can apply to several re-identification scenarios,including the following:– A theft in a crowded mall is observed by different people who give partial information aboutthe thief’s appearance. Based on this information, a first-pass search applies SBS methodsto cut down on the long surveillance video recordings from several cameras.– A mother has lost her child and can describe traits like clothes color and height of the child.Video surveillance material can be pruned and resulting suggestions can be displayed to themother.The above cases support the applicability of SBSs, but also reveal that together with the benefitsof such systems, come considerable risks such as that of erroneously pruning out the target of thesearch. This brings to the fore the need to jointly analyze the gains and risks of such systems.In the setting of human identification, we consider the scenario where we search for a specificsubject of interest, denoted as v ′ , belonging to a large and randomly drawn authentication groupv of n subjects, where each subject belongs to one of ρ categories. The elements of the set(authentication group) v are derived randomly from a larger population, which adheres to a setof population statistics. A category corresponds to subjects who adhere to a specific combinationof soft biometric characteristics, so for example one may consider a category consisting of blond,tall, females. We note the analogy to the scenario from chapter 3, but proceed to elaborate on thedifferent goal of the current chapter.With n being potentially large, we seek to simplify the search for subject v ′ within v by algorithmicpruning based on categorization, i.e., by first identifying the subjects that potentiallybelong to the same category as v ′ , and by then pruning out all other subjects that have not beenestimated to share the same traits as v ′ . Pruning is then expected to be followed by careful searchof the remaining unpruned set. Such categorization-based pruning allows for a search speedupthrough a reduction in the search space, from v to some smaller and easier to handle set S whichis the subset of v that remains after pruning, see Figure 4.1 and Figure 4.4. This reduction thoughhappens in the presence of a set of categorization error probabilities {ɛ f }, called confusion probabilities,that essentially describe how easy it is for categories to be confused, hence also describingthe probability that the estimation algorithm erroneously prunes out the subject of interest, byfalsely categorizing it. This confusion set, together with the set of population statistics {p f } ρ f=1which describes how common a certain category is inside the large population, jointly define thestatistical performance of the search pruning, which we will explore. The above aspects will beprecisely described later on.Example 7 An example of a sufficiently large population includes the inhabitants of a certaincity, and an example of a randomly chosen authentication group (n-tuple) v includes the set ofpeople captured by a video surveillance system in the aforementioned city between 11:00 and11:05 yesterday. An example SBS could be able to classify 5 instances of hair color, 6 instances ofheight and 2 of gender, thus being able to differentiate betweenρ = 5·6·2 = 60 distinct categories.An example search could seek for a subject that was described to belong to the first category of,say, blond and tall females. The subject and the rest of the authentication group of n = 1000people, were captured by a video-surveillance system at approximately the same time and placesomewhere in the city. In this city, each SBS-based category appears with probability p 1 ,··· ,p 60 ,and each such category can be confused for the first category with probability ɛ 2 ,··· ,ɛ 60 . The
45Figure 4.1: System overview.SBS makes an error wheneverv ′ is pruned out, thus it allows for reliability ofɛ 1 . To clarify, havingp 1 = 0.1 implies that approximately one in ten city inhabitants are blond-tall-females, and havingɛ 2 = 0.05 means that the system (its feature estimation algorithms) tends to confuse the secondcategory for the first category with probability equal to 0.05.What becomes apparent though is that a more aggressive pruning of subjects in v results in asmaller S and a higher pruning gain, but as categorization entails estimation errors, such a gaincould come at the risk of erroneously pruning out the subject v ′ that we are searching for, thusreducing the system reliability.Reliability and pruning gain are naturally affected by, among other things, the distinctivenessand differentiability of the subjectv ′ from the rest of the people in the specific authentication groupv over which pruning will take place that particular instance. In several scenarios though, thisdistinctiveness changes randomly because v itself changes randomly. This introduces a stochasticenvironment. In this case, depending on the instance in which v ′ and its surroundings v−v ′ werecaptured by the system, some instances would have v consist of bystanders that look similar tothe subject of interest v ′ , and other instances would havevconsist of people who look sufficientlydifferent from the subject. Naturally the first case is generally expected to allow for a lowerpruning gain than the second case.The pruning gain and reliability behavior can also be affected by the system design. At oneextreme we find a very conservative system that prunes out a member of v only if it is highlyconfident about its estimation and categorization, in which case the system yields maximal reliability(near-zero error probability) but with a much reduced pruning gain. At the other extreme,we find an effective but unreliable system which aggressively prunes out subjects in v, resultingin a potentially much reduced search space (|S|
- Page 1: FACIAL SOFT BIOMETRICSMETHODS, APPL
- Page 5: AcknowledgementsThis thesis would n
- Page 8: 6hair, skin and clothes. The propos
- Page 11 and 12: 97 Practical implementation of soft
- Page 13 and 14: 11Notations used in this workE : st
- Page 15 and 16: 13Chapter 1IntroductionTraditional
- Page 17 and 18: 15event of collision, which is of s
- Page 19 and 20: 17ric. In Section 6.6 we employ the
- Page 21 and 22: 19Chapter 2Soft biometrics: charact
- Page 23 and 24: 21is the fusion of soft biometrics
- Page 25 and 26: 23plied on low resolution grey scal
- Page 27 and 28: 25Chapter 3Bag of facial soft biome
- Page 29 and 30: 27In this setting we clearly assign
- Page 31 and 32: 29Table 3.1: SBSs with symmetric tr
- Page 33 and 34: 31corresponding to p(n,ρ). Towards
- Page 35 and 36: the same category (all subjects in
- Page 37 and 38: 3.5.2 Analysis of interference patt
- Page 39 and 40: an SBS by increasing ρ, then what
- Page 41 and 42: 39Table 3.4: Example for a heuristi
- Page 43 and 44: 41for a given randomly chosen authe
- Page 45: 43Chapter 4Search pruning in video
- Page 49 and 50: 472.52rate of decay of P(τ)1.510.5
- Page 51 and 52: 49to be the probability that the al
- Page 53 and 54: 51The following lemma describes the
- Page 55 and 56: 534.5.1 Typical behavior: average g
- Page 57 and 58: 55n = 50 subjects, out of which we
- Page 59 and 60: 5710.950.9pruning Gain r(vt)0.850.8
- Page 61 and 62: 59for one person, for trait t, t =
- Page 63 and 64: 61Chapter 5Frontal-to-side person r
- Page 65 and 66: 63Figure 5.1: Frontal / gallery and
- Page 67 and 68: 6510.90.80.7Skin colorHair colorShi
- Page 69 and 70: 6710.90.80.70.6Perr0.50.40.30.20.10
- Page 71 and 72: 69Chapter 6Soft biometrics for quan
- Page 73 and 74: 71raphy considerations include [BSS
- Page 75 and 76: 73Figure 6.3: Example image of the
- Page 77 and 78: 75A direct way to find a relationsh
- Page 79 and 80: 77- Pearson’s correlation coeffic
- Page 81 and 82: 79shown to have a high impact on ou
- Page 83 and 84: 81Chapter 7Practical implementation
- Page 85 and 86: 834) Eye glasses detection: Towards
- Page 87 and 88: 857.2 Eye color as a soft biometric
- Page 89 and 90: 87Table 7.5: GMM eye color results
- Page 91 and 92: 89and office lights, daylight, flas
- Page 93 and 94: 917.5 SummaryThis chapter presented
- Page 95 and 96: 93Chapter 8User acceptance study re
44 4. SEARCH PRUNING IN VIDEO SURVEILLANCE SYSTEMSfurther elimination <strong>of</strong> categories, which limits large databases <strong>of</strong> subjects to a fraction <strong>of</strong> the initialdatabase, see Figure 4.1. In the context <strong>of</strong> this chapter the elimination or filtering <strong>of</strong> the employedcategories is based on the s<strong>of</strong>t biometric characteristics <strong>of</strong> the subjects. The pruned database canbe subsequently processed by humans or by a biometric such as face recognition.The approach <strong>of</strong> pruning the search using SBSs, can apply to several re-identification scenarios,including the following:– A theft in a crowded mall is observed by different people who give partial information aboutthe thief’s appearance. Based on this information, a first-pass search applies SBS methodsto cut down on the long surveillance video recordings from several cameras.– A mother has lost her child and can describe traits like clothes color and height <strong>of</strong> the child.Video surveillance material can be pruned and resulting suggestions can be displayed to themother.The above cases support the applicability <strong>of</strong> SBSs, but also reveal that together with the benefits<strong>of</strong> such systems, come considerable risks such as that <strong>of</strong> erroneously pruning out the target <strong>of</strong> thesearch. This brings to the fore the need to jointly analyze the gains and risks <strong>of</strong> such systems.In the setting <strong>of</strong> human identification, we consider the scenario where we search for a specificsubject <strong>of</strong> interest, denoted as v ′ , belonging to a large and randomly drawn authentication groupv <strong>of</strong> n subjects, where each subject belongs to one <strong>of</strong> ρ categories. The elements <strong>of</strong> the set(authentication group) v are derived randomly from a larger population, which adheres to a set<strong>of</strong> population statistics. A category corresponds to subjects who adhere to a specific combination<strong>of</strong> s<strong>of</strong>t biometric characteristics, so for example one may consider a category consisting <strong>of</strong> blond,tall, females. We note the analogy to the scenario from chapter 3, but proceed to elaborate on thedifferent goal <strong>of</strong> the current chapter.With n being potentially large, we seek to simplify the search for subject v ′ within v by algorithmicpruning based on categorization, i.e., by first identifying the subjects that potentiallybelong to the same category as v ′ , and by then pruning out all other subjects that have not beenestimated to share the same traits as v ′ . Pruning is then expected to be followed by careful search<strong>of</strong> the remaining unpruned set. Such categorization-based pruning allows for a search speedupthrough a reduction in the search space, from v to some smaller and easier to handle set S whichis the subset <strong>of</strong> v that remains after pruning, see Figure 4.1 and Figure 4.4. This reduction thoughhappens in the presence <strong>of</strong> a set <strong>of</strong> categorization error probabilities {ɛ f }, called confusion probabilities,that essentially describe how easy it is for categories to be confused, hence also describingthe probability that the estimation algorithm erroneously prunes out the subject <strong>of</strong> interest, byfalsely categorizing it. This confusion set, together with the set <strong>of</strong> population statistics {p f } ρ f=1which describes how common a certain category is inside the large population, jointly define thestatistical performance <strong>of</strong> the search pruning, which we will explore. The above aspects will beprecisely described later on.Example 7 An example <strong>of</strong> a sufficiently large population includes the inhabitants <strong>of</strong> a certaincity, and an example <strong>of</strong> a randomly chosen authentication group (n-tuple) v includes the set <strong>of</strong>people captured by a video surveillance system in the aforementioned city between 11:00 and11:05 yesterday. An example SBS could be able to classify 5 instances <strong>of</strong> hair color, 6 instances <strong>of</strong>height and 2 <strong>of</strong> gender, thus being able to differentiate betweenρ = 5·6·2 = 60 distinct categories.An example search could seek for a subject that was described to belong to the first category <strong>of</strong>,say, blond and tall females. The subject and the rest <strong>of</strong> the authentication group <strong>of</strong> n = 1000people, were captured by a video-surveillance system at approximately the same time and placesomewhere in the city. In this city, each SBS-based category appears with probability p 1 ,··· ,p 60 ,and each such category can be confused for the first category with probability ɛ 2 ,··· ,ɛ 60 . The