AB STRACT In t his paper, we describe a privacy filt er proposed for t he Visual Privacy T ask (VPT ) 2013, as a case st udy t o validat e t he efficacy of t he User-Int imat e Requirement s and Evaluat ion Met hodology (UI-REF). T his comprises an aut omat ed object ive evaluat ion phase support ed by a subject ive user st udy t o crossvalidat e and complement t he result s of t he object ive assessment . T he performance of t he proposed filt er, as well as t he overall performance t rends and t radeoffs of t he alt ernat ive filt ering t echniques are highlight ed. T he result s demonst rat ed t he consist ency and high resolut ion evaluat ion capabilit ies of t he Holist ic Evaluat ion Framework for Privacy by Co-Design and Impact Assessment for benchmarking privacy filt ering solut ions.
A variet y of image processing t echniques have been proposed t o mit igat e privacy prot ect ion failure risks. T hese video privacy filt ering approaches have essent ially applied filt ering t echniques t o obscure t he privacy-sensit ive part s of t he capt ured video; in a similar way as t he est ablished pract ice in t he film and t elevision sect or. T he t rade-off bet ween t he levels of masking, and, t he informat iveness of a video image means t hat naïve deployment of such privacy filt ers could lead t o video surveillance syst ems t hat are pot ent ially ineffect ive in eit her adequat ely prot ect ing t he privacy of t he cit izen or in ret aining t he essent ial informat ion for t he int ended securit y monit oring in t he given sit uat ion. T hus, syst emat ic and comprehensive evaluat ion of t he privacy filt ering solut ions is necessary. Many report ed at t empt s at evaluat ing t he performance and impact of privacy filt ering have adopt ed a relat ively limit ed analysis perspect ive. T he VPT 2013 t ask evaluat ion met hodology has responded t o t he need for a more inclusive, holist ic and high resolut ion assessment of privacy filt ering requirement s as well as t he evaluat ion of t he efficacy and impact s of t he result ing privacy filt ering solut ions based on t he UI-REF met hodology [1]. T he PEViD dat aset [2] was used for t he impact assessment of alt ernat ive privacy prot ect ion solut ions.
As a case st udy, we proposed a privacy filt er t hat could be applied aut omat ically t o CCT V video feed. T his aimed: i ) t o preserve maximum privacy (anonymit y) by applying t he filt er t o t he foreground regions including t he human body, and, i i ) t o ret ain t he non-privacy-sensitive informat ion so as t o deliver some surveillance value (int elligibilit y); i i i ) t o minimise t he pot ent ial viewer’s dist ract ion and annoyance caused by t he deployed privacy filter. T he sect ions below describe t he st eps.
T o meet t he first requirement , t he privacy filt er was applied t o t he subject region wit hin t he scene. A st at e-of-t he-art object segment at ion algorit hm as described in [3] was used in order t o obt ain t he init ial foreground mask followed by morphological operat ions t o smoot h it furt her. T he generat ed mask has been used as a basis for privacy filt er applicat ions.
T o hide t he det ails of t he appearance of t he viewed subject , a
Discret e Cosine T ransform (DCT ) was applied over each 8x8
sub-block of t he subject bounding box. T hen t he sign was flipped
for four (
A Sobel filt er was applied t o t he human figure t o out line t he st rong edges using t he saturat ion component of t he HSV colour space. T he sat urat ion channel was chosen as it yields good cont rast under all light ing condit ions as well as preserving t he out line edges. T he edges were t hresholded and at t enuat ed by modulat ing t heir pixel value by t he inverse of t he squared dist ance t o t he cent re, keeping only t he st rongest cent ral ones. T he result s of t he scrambling and edge det ect ion were applied t o t he subject mask area and t hen blended int o t he original image wit h a radial at t enuat ion so as t o prevent t he occurrence of highly visible and irrit at ing edges around t he filt ered areas of t he video. Figure 1 shows t he out put of t he proposed approach.
T he object ive and subject ive evaluat ion crit eria were based on t he UI-REF Co-design, Evaluat ion and Impact Assessment Met hodology [1, 4, 5]. A subset of t he Key Performance Indicat ors (KPIs) was deployed from: Efficacy, Consist ency, Int elligibility, Disambiguit y, Aest het ics (st ruct ural, t ext ural, t onal, harmony and symmet ry maps), Select ivit y, Sensit ivit y, Comput at ional efficiency, real-t ime Web-Scalabilit y and Vulnerabilit y t o at t ack. Subject ive evaluat ion was based on UIREF Qualit y-of-Experience, Effect s, Side-Effect s and Affect s.
T he object ive crit eria to assess t he opt imalit y of t he balance of
privacy prot ect ion and securit y assurance as may be offered by a
given privacy filt ering solut ion were defined as follows:
(
A user st udy was conduct ed t o complement and cross-validat e
t he result s of the object ive t est . T he quest ionnaire probed for
t he Likert -Scale based assessment of t he same KPIs as t he above.
(
T his sect ion provides t he evaluat ion result s of t he proposed
filt er plus an overview of t he performance t rends in VPT 2013.
T he report ed result s of t he object ive t est are t he average score
of eleven (11) videos which were object ively evaluat ed in t erms
of t he defined crit eria. Figure 2, shows t he performance of t he
proposed filt er as markedly above t he average of all VPT 2013
submit t ed solut ions for t he Intelligibility and Appropriateness
crit eria. As for Privacy, t he proposed filt er performed below t he
average score. T his was part ially because a few submissions were
st rong filt ers in which t he ot her crit eria were compromised t o
achieve higher privacy. However, t he overall t rends across all
t he alt ernat ive privacy filt ering solut ions emphasised t he t
radeoff bet ween t he Intelligibility and Privacy as well as showed a
direct relat ionship bet ween Intelligibility and Appropriateness.
T he subject ive st udy involved over sixt y (60) part icipant s who
were equally dist ribut ed t o examine five (
T his paper has proposed a video privacy filt er using a t ransform domain scrambling met hod and edge det ect ion; t o achieve t he highest privacy wit h minimum cont ent dist ort ion and viewer dist ract ion. T he UI-REF based Holist ic Evaluat ion Framework has deployed a sub-set of it s KPIs t o provide an int egrat ive and consist ent set of object ive and subject ive assessment crit eria wit h demonst rat ed consist ency. T his has built on a similar approach as for VPT 2012; t he successful deployment of t his framework mot ivat es furt her innovat ion in assessment met rics for evaluat ion of privacy-filt ering and risk mit igat ion t echnologies.
T his work was support ed by t he European Commission under cont ract s FP7-261743 VideoSense project .