=Paper=
{{Paper
|id=Vol-2540/paper32
|storemode=property
|title=None
|pdfUrl=https://ceur-ws.org/Vol-2540/FAIR2019_paper_56.pdf
|volume=Vol-2540
}}
==None==
https://ceur-ws.org/Vol-2540/FAIR2019_paper_56.pdf
Human activity recognition using deep learning
approaches
Bradley J. Pillay 1 [0000-0002-0369-4680], Anban Pillay1 [0000-0001-7160-6972], and Edgar Jem-
bere1 [0000-0003-1776-1925]
1 School of Mathematics, Statistics and Computer Science, University of KwaZulu-Natal,
Westville Campus, Private Bag X54001, Durban 4000, South Africa
215031687@stu.ukzn.ac.za
Abstract. Human activity recognition using video data has been an active re-
search area in computer vision for many years. This research presents an archi-
tecture that employs deep learning techniques to effectively solve human activi-
ty recognition of single individuals using visual information from videos. The
architecture adopts an Octave Convolutional neural network as a feature extrac-
tor and a weighting strategy that regresses the importance of each video seg-
ment. This architecture was trained and evaluated on the KTH human activity
dataset. The results are promising and validates the approach taken.
Keywords: human activity recognition (HAR), octave convolution neural net-
work, temporal segment network, deep adaptive temporal pooling
1 Introduction
Human activity recognition is the process of identifying the actions and goals of indi-
viduals from a series of observations of activities performed within a given environ-
ment. This recognition process can be applied to many areas that have the goal of
monitoring human actions such as surveillance systems for detecting illegal activities
or sporting arenas for detecting foul play. This recognition task is solvable using ei-
ther video information, sensor data, or both.
The application of deep learning has shown tremendous performance improve-
ments over traditional techniques. The use of deep learning has an advantage through
its automatic feature extraction abilities [1]. In this paper, a deep learning architecture
that can successfully classify human activities is presented. The performance of the
architecture on classifying activities of the KTH human activity dataset is also given.
2 Proposed Architecture Design
The proposed architecture utilizes a temporal segment network (TSN) as the base
architecture. A video V is divided into N segments of equal lengths. A frame from the
middle of each video segment serves as input for the spatial stream. The motion rep-
resentations of each segment were also extracted using the TVL1 optical flow algo-
Copyright © 2019 for this paper by its authors. Use permitted under Creative Commons License
Attribution 4.0 International (CC BY 4.0)
2
rithm on a stack of 10 consecutive frames. An OctResNet50 model pre-trained on the
CIFAR-100 dataset was used as the frame-level feature extractor. The OctResNet50
model is a ResNet50 model that utilizes the octave convolution [4]. The temporal
features are stacked together and parsed to a deep adaptive temporal pooling (DATP)
module to generate the importance of each temporal segment [2]. The two-component
Gaussian Mixture Model (GMM) was chosen as the weights-generator employed by
the DATP module. The generated weights were assigned to the temporal segments
and a late fusion strategy was adopted for fusing both the spatial and temporal
streams. The fused vector was then parsed to a feedforward neural network that out-
puts the predicted class label for the entire video. All trainable parameters in this ar-
chitecture was trained through backpropagation.
3 Results
Table 1 displays the confusion matrix obtained by the proposed architecture on the
test set made up of 121 videos from the KTH dataset [5]. The accuracy rate obtained
by the proposed architecture is 63.6%. The state-of-the-art results of 95.33% classifi-
cation accuracy were achieved in [3].
Table 1. Confusion Matrix obtained by the proposed architecture on the KTH dataset
Actual Class Predicted Class Labels
Labels Boxing Handclapping Handwaving Jogging Running Walking
Boxing 18 0 1 0 1 0
Handclapping 1 18 1 0 1 0
Handwaving 1 2 17 0 0 0
Jogging 1 1 0 5 10 4
Running 0 1 0 1 14 4
Walking 0 0 0 5 10 5
4 Conclusion
Most of the boxing, handclapping and handwaving videos were classified correctly.
The jogging and walking actions were misclassified as running. This misclassification
may have occurred because the motion representations of the 3 actions were similar.
The results obtained are not as good as the state-of-the-art results, but they are prom-
ising and validates the approach taken.
References
1. Khurana, R. and Kushwaha, A.K.S., 2019. Delving Deeper with Dual-Stream CNN for
Activity Recognition. In Recent Trends in Communication, Computing, and Electronics
(pp. 333-342). Springer, Singapore.
2. Song, S., Cheung, N.M., Chandrasekhar, V. and Mandal, B., 2018. Deep Adaptive Tem-
poral Pooling for Activity Recognition. arXiv preprint arXiv:1808.07272.
3
3. Arunnehru, J., Chamundeeswari, G. and Bharathi, S.P., 2018. Human action recognition
using 3D convolutional neural networks with 3D motion cuboids in surveillance videos.
Procedia computer science, 133, pp.471-477.
4. Chen, Y., Fang, H., Xu, B., Yan, Z., Kalantidis, Y., Rohrbach, M., Yan, S. and Feng, J.,
2019. Drop an Octave: Reducing Spatial Redundancy in Convolutional Neural Networks
with Octave Convolution. arXiv preprint arXiv:1904.05049.
5. Nada.kth.se. (2019). Recognition of human actions. [online] Available at:
http://www.nada.kth.se/cvap/actions/ [Accessed 14 Aug. 2019]