Page 136 - Proceedings of the 2018 ITU Kaleidoscope

2018 ITU Kaleidoscope Academic Conference

increasing the number of clusters does not guarantee the block level and frame level accuracy are listed in Table 4
improvement of the system performance. and Table 5.

Table 2– Performance of the system with varied number of
clusters in UMN Dataset
Table 4– Comparison of works (Block level accuracy)

No of Performance Datasets
Clusters
Accuracy Recall Precision Method UMN UCSD Created
Ped 1 Ped 2 Dataset
4 98.94 98.68 100
HOFME
5 98.17 98.66 98.66 98.52 72.70 87.50 95.04
[12]
6 98.94 100 92.73
Proposed 98.94 71.32 88.13 98.78
7 98.78 98.68 96.10 Method

4.2.3. Block Size Table 5 – Comparison of works (Frame level accuracy)

Each frame in the video is divided into M × N uniform Datasets
blocks for computing influence weight between blocks and UCSD
further processing. However, considering that the motion Method UMN Created
influence value represents the motion of surrounding blocks Dataset
rather than the target block itself, the choice of the block Ped 1 Ped 2
size can affect the performance. Therefore, we measured HOFME
the unusual activity detection performance for the UMN [12] 84.94 86.30 89.50 93.56
dataset on various block sizes to show the effect of a
change in block size. Proposed
Method 92.35 81.20 91.10 95.60
The experimental results considering different block size is
shown in Table 3. When the size of a block is small and the 4. CONCLUSION
motion information of an object is represented by various
motion vectors, thereby generating noise, the use of a Abnormal activity detection in a crowded scene requires
motion influence map will be fully considered. However, if development of system model and associated learning
the size of a block is bigger than the size of an object, process. Unlike previous methods described in the literature,
important motion characteristics may be disregarded. Here, which have focused on either local or global abnormal
it can also be observed that the performance is maximized activity detection, the proposed method considers the
when the block size is set to approximately half the size of motion characteristics within a frame to detect and localize
the pedestrian. abnormal human activities in a crowded scene. The model
can classify a frame as normal, abnormal, and localize the
Table 3– Performance of the system with varied block areas of abnormal activities within the frame. The
division of frames in UMN Dataset experiments were conducted on two public datasets, UMN
and UCSD datasets to validate the effectiveness of the
Frame Performance proposed method in comparison with competing methods
Division Accuracy Recall Precision from the literature. The proposed solution was also tested
with a custom made dataset representing a local real time
8×6 98.94 98.68 100 environment. The ITU recommendations were used in
developing system in a standardized modular fashion.
10 × 8 96.35 97.20 99.25 However, one of the limitation of the proposed system is:
threshold of acceptance need to be fixed for specific
4.3. Comparison with other works scenarios. Another constraint of the model is that it is more
dependent on view angle and camera distance from action.
The proposed algorithm has been compared with HOFME The experiments were limited to a fixed viewpoint, and
[12] and has produced higher accuracy in both standard there is a limitation in the applicability of the approach for
datasets (UMN, UCSD) and custom made dataset. Both surveillance cameras with zoom, or tilt functionality.

– 120 –