IEEE Consumer Electronics Magazine - July 2019 - 35

the n k map, the weighted variance generates the v k map by measuring the pixelwise variance between all three frames, which are
also weighted differently according to c k . Each v k (x, y) that
exceeds the p level indicates a positive OV ( | k (x, y) = 1) .

DATA FILTERING
Before the elements of the | k map are analyzed for any
abnormal event, each OV is normalized against the camera's
perspective by using normalization weights:
? | k = / | k (x, y) $ B (x, y) $ N (x, y) .

(3)

x, y

A moving median filter could be further applied on the ? | k
map to remove any unwanted OV signals. Because OVs in
? | k are representing object motion instead of object acceleration, the ? | k map is further processed by calculating its first
derivative to obtain the object acceleration ` ? | k j .

UNUSUAL EVENT DETECTION
The detector measures the consistency of ? | k as
i = tan -1

? | k +9 - ? | k
.
5$w$h

(4)

The unusual event is considered when i 2 1 over the observational period; otherwise, the candidate frame is ignored. The
system will reset when | k is empty.

Table 1. The standards of classification reliability
of the proposed unusual event detection on an
original data set.
Video

ACC

PRC

SEN

SPE

Latency (s)

1

0.984

1.000

0.967

1.000

0.433

2

0.977

1.000

0.967

1.000

0.167

3

0.994

1.000

0.993

1.000

0.100

4

0.985

0.976

0.992

0.978

0.167

5

0.985

1.000

0.935

1.000

0.233

6

0.952

1.000

0.667

1.000

0.500

7

0.994

1.000

0.986

1.000

0.600

8

0.904

0.851

1.000

0.787

0.200

9

0.994

1.000

0.886

1.000

1.367

10

0.999

1.000

0.981

1.000

0.100

11

0.997

1.000

0.942

1.000

0.433

12

1.000

1.000

1.000

1.000

N/A

13

1.000

1.000

1.000

1.000

N/A

14

1.000

1.000

1.000

1.000

N/A

Note: Latency is measured on videos that contain true positive
detection. N/A: not applicable.

With more comprehensive features
to process, BMM provides fewer
false positives; however, it requires
a training process and high
complexity.
AN IMPLEMENTATION FOR CASE STUDY
The proposed method was first assessed on 14 actual surveillance footages with different fleeing en masse scenarios consisting of normal and/or abnormal events, distinct lighting
(dark or bright environment), unique scenes (indoor or outdoor), and various camera perspectives. The ground truths
for these videos were generated manually. Second, the performance of prior arts (including AMC [5], MHOF [8],
BMM [9], MAD [6] and ELD [7]) were benchmarked
against the proposed system on a publicly available University of Minnesota (UMN) crowd data set [3] that comprises
11 videos with fleeing en masse scenarios for each video on
three different scenes. The ground truths are available from
within the data set. The results had been compared framewise against the ground truth and are statistically analyzed in
terms of accuracy (ACC), precision (PRE), true positive rate
(SEN), and true negative rate (SPE). The range of scores is
0 (worst) to 1 (best).

DETECTION PERFORMANCE ON AN
ORIGINAL DATA SET
Table 1 shows the comprehensive assessment of the proposed
method based on the original data set. From the results, it is

(a)

(b)

FIGURE 3. The detection method observes the OVs from the
input video by frame and issues a user-based alert when a certain
number of OVs is evident (e.g., k = 1, 689th of video 7).
(a) k = 700th. (b) k = 1, 689th.

JULY 2019

^

IEEE CONSUMER ELECTRONICS MAGAZINE

35
IEEE Consumer Electronics Magazine - July 2019

Table of Contents for the Digital Edition of IEEE Consumer Electronics Magazine - July 2019
