Using intersubject correlation of fMRI data to ...

6 downloads 0 Views 952KB Size Report
Introduction. Results. A recent study that compared CCTV operators to novices when performing the task of judging harmful intent from actual CCTV.
Recognizing Harmful Intent from Surveillance Video Viewed Through the Eye-movements of Novice and Experienced Observers Frank E

1 Pollick ,

Greta

1 Todorova ,

Steven

2 Thurman ,

Joseph

2 Burling ,

Hongjing

2 Lu

1School

of Psychology, University of Glasgow, UK 2Psychology Department, University of California, Los Angeles

Results

Methods & Procedure

A recent study that compared CCTV operators to novices when performing the task of judging harmful intent from actual CCTV video showed differences in brain activity and ability to recognize harmful intent (Petrini McAleer, Neary, Gillard & Pollick, 2014). Subsequent research has revealed statistical differences in the eye movements of CCTV operators and novices (Roffo, Cristani, Pollick, Segalin & Murino, 2013) and that experienced operators appear more consistent at tracking important features (Burling, Lu, Todorova & Pollick, 2016). It is known that “seeing” the visual world through the gaze pattern of experts can be used to train performance on tasks involving visually complex scenarios (Wilson et al., 2011; Vine et al., 2013). Here we conducted an experiment to examine whether seeing the videos through the gaze patterns of either a novice or an experienced operator would influence performance at recognising harmful intent.

• Participants were 44 students with no CCTV experience. • Each participant viewed a random ordering of 24 videos, 6 from each category (fight’, ‘confrontation’, ‘play’, ‘nothing’) with 3 videos per category derived from an operator’s gaze and 3 from a novice’s gaze. The selection of which 3 videos and which operator or novice’s gaze pattern was also randomised. Designed against idiosyncratic aspects of video or gaze path influencing results • After presentation of a video, two questions asked: • ‘Did a violent incident occur after the end of the video?’ • ‘How confident are you?’ • Data were analysed using a signal detection analysis, where recognising a fight video as ending in violence was coded as a hit • A within subjects ANOVA was used to analyse the participants’ confidence ratings of different video categories.

Stimuli

Results

The study of Petrini et al. (2014) also obtained eye tracking data of novices and experienced CCTV operators. We used this data to filter the videos so that the central location of the gaze would appear at the original resolution, while the surrounding area would be blurred (Geisler, Perry & Najemnik, 2006). This produced two sets of videos: one created with Operator gaze patterns and another with Novice gaze patterns.

Frames from a processed demonstration video

Operator Gaze Pattern Operator

3.4 3 2.6 2.2 1.8 1.4 1 0.6 0.2 -0.2 -0.6 -1

Novice Gaze Pattern Novice

1.35

*

*

6 *

5 4 3 2

4.56

5.14

4.23

4.40

1 0 Confrontation

Fight Nothing Video Category

Play

• Minimal difference in sensitivity and bias between videos revealing Operator and Novice gaze patterns • Trend for higher sensitivity and less bias when viewed through Novice gaze patterns • Confidence greater in judging ‘Nothing’ than all other displays, also significant difference between ‘Confrontation’ and ‘Fight’ • Current results using randomised presentation order reveal no spontaneous benefit for novice observers to view actions through the gaze pattern of Operators; instead opposite trend found • Further work on the effects of training required – these results provide important baseline results

References

0.88 0.2 -0.03

d' (sensitivity)

7

Summary & Conclusions

Sensitivity and bias results for predicting a violent outcome

Mean sensitivity and bias indices

Our stimuli came from the set of Petrini et al. (2014). These included a total of 24 videos from 4 categories ( ‘fight’, ‘confrontation’, ‘play’, ‘nothing’) and each was 16-second in duration. Each video stopped before any physical violence took place. A screenshot of an anonymised video is shown to the right with the fixation location of several observers superimposed.

Confidence results for each video category

Mean Confidence Levels

Introduction

C (bias)

Burling, J.M., Lu, H., Todorova, G. & Pollick, F.E. (2016). A comparison of eye-movement patterns between experienced observers and novices in detecting harmful intention from surveillance video. VSS Annual Conference. Geisler, W. S., Perry, J. S., & Najemnik, J. (2006). Visual search: the role of peripheral information measured using gaze-contingent displays. Journal of Vision, 6(9), 858–73. Petrini, K., McAleer, P., Neary, C., Gillard, J., & Pollick, F. E. (2014). Experience in judging intent to harm modulates parahippocampal activity: an fMRI study with experienced CCTV operators. Cortex; a Journal Devoted to the Study of the Nervous System and Behavior, 57, 74–91. Roffo, G., Cristani, M., Pollick, F., Segalin, C., & Murino, V. (2013). Statistical analysis of visual attentional patterns for video surveillance. Lecture Notes in Computer Science, 8259 LNCS(PART 2), 520–527 Vine, S. J., Chaytor, R. J., McGrath, J. S., Masters, R. S. W., & Wilson, M. R. (2013). Gaze training improves the retention and transfer of laparoscopic technical skills in novices. Surgical Endoscopy, 27(9), 3205–3213. Wilson, M. R., Vine, S. J., Bright, E., Masters, R. S. W., Defriend, D., & McGrath, J. S. (2011). Gaze training enhances laparoscopic technical skill acquisition and multi-tasking performance: a randomized, controlled study. Surgical Endoscopy, 25(12), 3731–3739. This work is supported by NSF BSC 1353391 grant.

Suggest Documents