ADCAIJ: Advances in Distributed Computing and Artificial Intelligence Journal
Regular Issue, Vol. 13 (2024), e31638
eISSN: 2255-2863
DOI: https://doi.org/10.14201/adcaij.31638

Classification of Animal Behaviour Using Deep Learning Models

M. Sowmya^a, M. Balasubramanian^b and K.Vaidehi^c

^a Research Scholar, Department of CSE, Annamalai University, Annamalai Nagar, India

^b Associate Professor, Department of CSE, Annamalai University, Annamalai Nagar, India

^c Associate Professor, Department of CSE, Stanley College of Engineering and Technology for Women, Hyderabad, India

✉ msoumya@stanley.edu.in

ABSTRACT

KEYWORDS

1. Introduction

The term behaviour is used by animal scientists to describe what an animal does during daily life. Animal recognition or wildlife has been an area of great interest to biologists. Given that there are distinct categories of «animals» manually recognizing them can be a difficult task. Areas such as farming, railway monitoring, highway monitoring, national parks and reserved forests require the extensive surveillance of the animals that cause issues such as crop destruction, road blockage etc. Animal classification systems are beneficial for research connected to behavioural training focused on animals. In addition, harmful animal disturbance in the domestic area can be avoided through the use of such systems. Thus, a deep learning algorithm that classifies animals on the basis of their images can help «monitor» them more efficiently. Classifying images of wildlife through computer technology can help people identify wildlife, which is of great significance to help people understand and protect it. Animal classification plays a very distinguished role in multiple professions and activities such as farmers, railway track monitoring, highway monitoring, etc. It is a difficult task for a person to monitor and classify animals for many hours. An image classifier becomes useful in such situations as it automates the process of identifying and classifying animals. Agriculture is the backbone of the Indian economy, where more than 60 % of the country’s population is directly or indirectly dependent on this sector. As the global population is growing continuously it is necessary to significantly increase food production. The food needs to have a high nutritional value and its security must be guaranteed around the world. Crop damage by animal intrusion is one of the major threats in the loss of crop yield. Productivity is decreased when wild animals tramp over harvests and eat them. Computer vision techniques have the potential to provide security from wild animals in agriculture. Any animal’s sound can be helpful to human beings in terms of greater security or as a means of predicting natural disasters. Most elephant-train collisions occur due to a lack of adequate reaction time due to poor driver visibility at sharp turns, night-time operation, and poor weather conditions. Furthermore, railway services incur significant financial losses and disruptions to services annually, due to such accidents.

2. Related Work

The classification of the images of animal species was used by Prudhivi et al. (2023) to classify animals in real time in forests. The dark net algorithm was employed for wildlife surveillance by Manasa (2021); this algorithm , had a pre-trained dataset in the YOLOv3 model. Bimantoro et al. (2021) used a «convolutional neural network» to minimize the workload of sheep farmers. The behaviour of dairy cows, such as (drinking and ,walking, is linked to their physiological well-being; a CNN-LSTM was used to assess these behaviours (Wu et al., 2021). Zeng et al. (2021) focused on the classification of comparable animal photos using a basic 2D CNN implemented in Python. The authors focused on snub-nosed monkeys and regular monkeys as binary classifications. Deep CNN with genetic segmentation was used by Chandrakar et al. (2021), as a method for the autonomous detection and recognition of animals. One of the biggest issues associated with in intelligent livestock management is achieving a good accuracy in the identification of the breed of domestic animals using photographs (Ghosh et al., 2021). Ecological video traps are a popular method of monitoring an ecosystem animal population since they provide continuous information without being obtrusive. Monitoring animal behaviour outdoors necessitates the use of accurate techniques for identifying species, which mostly rely on image data recorded by camera traps ( Favorskaya et al., 2019). Recent studies have dedicated little attention the detection of large animals in photographs containing road scenes. The development of such a data set using Google Open Images and COCO datasets was described by Yudin et al. (2019). Moreover, the implementation of the «real-time animal vehicle» vehicle collision mitigation system was proposed in the same paper using a deep learning approach. The success of classification systems was determined by Neena et al. (2018) using feature extraction methods. Chen et al. (2017) presented a deep neural fish classification system that used a camera to automatically name fish without human intervention; the photos of marine animals were collected by an underwater robot with an implanted gadget. In a research study by Bhavani et al. (2019), convolutional neural networks on Android were used to identify dog breeds. The use of neural networks for mobile terminals is becoming more common as the volume of picture data processed by mobile devices grows. MobileNetV3 achieved a superior balance of efficiency and accuracy for real-world picture classification tasks on mobile terminals, as demonstrated by Qian et al. (2021). In the study of Vehkaoja et al. (2022), data from 45 middle to large dogs was obtained from seven static and dynamic dog activities (sitting, standing, searching i.e., sniffing, etc.), with six degrees of freedom movement sensors attached to the collar and harness. The data collection method was repeated with 17 dogs. A deep learning framework for tracking and categorizing the five common behaviours of cows, namely, feeding, exploring, grooming, walking and standing, was proposed by (Qiao et al., 2022). The study of Brandes et al. (2021) involved three captive giraffes to test two distinct commercially available high-resolution accelerometers, e-obs and Africa wildlife «tracking» (AWT), and analysed the accuracy of automatic behaviour classifications, focusing on the «random» forests algorithm. The YOLOv3 method was used in the proposal of Deng et al. (2021) to detect the behaviour and posture of sheep. Average precision in this study was 92.47 percent. A review of machine learning algorithms for detecting farm animal behaviours such as lameness, grazing, rumination etc., was carried out by Debauche et al. (2021). The goal of Fogarty et al. (2020) was to explore feature creation and ML algorithm possibilities so as to provide the most precise behavioural characterization of extensively grazed sheep via an ear-borne accelerometer. The study of Sakai et al. (2019) used machine learning algorithms to classify goat behaviours using a back-mounted 9-axis multi sensor and changes in the predictive scores were evaluated by equalizing the prevalence of each behaviour. Individual monitoring using barcodes has revolutionized the study of animal behaviours. Individual animal behaviour and social dynamics that characterize populations and species were studied through animal monitoring by Ferrarini et al. (2022). Kamminga et al. (2017) presented a system for embedded platforms that uses multitask learning to address these issues. Many farms have used video surveillance technology, which has resulted in a vast volume of video data (Yang et al., 2020). Traditional image processing methods and deep learning approaches were introduced, as well as their applications. Williams et al. (2017) explored if the head-neck position and activity of cattle can be used to track drinking as well as whether triaxial accelerometers worn around the neck may detect drinking. Wang et al. (2020) provided a method for classifying wild animal photographs based on transfer - learning. Nguyen et al. (2017) offered a framework for developing automated animal recognition in the outdoors, with the goal of creating an automated wildlife monitoring system. Billah et al. (2022) proposed a deep learning strategy designated to create a completely automated pipeline for goat face identification and recognition.

3. Proposed Methodology

The goal of this work is to develop a means of classifying animal behaviours using CNN and Resnet models. The proposed system detects the behaviours of animals such as eating, sitting and standing using animal video frames. The main steps involved in this work are as follows: dataset collection, converting videos to frames, training the 2D-CNN model, VGG16 and ResNet50 models, followed by model testing the block diagram of the proposed system is given in Figure. 1.

3.1. 2D-CNN Architecture

The detailed description of each layer is given as:

1) Input layer

An input layer, an output layer, and a large number of hidden layers make up a convolutional neural network. Image and video recognition, image classification, medical image analysis, computer vision, and natural language processing are applications of CNN. The colour image has 3 (RGB) colour channels, and the grayscale image depth is one. The stride refers to the number of steps in which the filter slides over the input. When the stride value is 1, the filters are moved one pixel at a time. When the stride value is 2, they move 2 pixels at a time. Padding allows to alter the size of the output. When convolution is applied to an input, the produced output size of the matrix is reduced, resulting in information loss. To avoid this, the padding concept has been implemented. Padding is done through the input volume with zeros at the border. Valid and same are two popular padding options. The same padding indicates that the output size remains the same as the input size, and valid padding means no padding is added. The convolution operation is shown in equation (1) and the feature map size is shown in equation (2) (Sowmya et al., 2023), 2D CNN architecture is shown in Figure 2 and convolution operationis shown in Figure 3.

$${\mathrm{y}}_{\text{conv }}=\mathrm{f}(\underset{j=0}{\overset{j=1}{\Sigma }}\underset{\mathrm{i}=0}{\overset{\mathrm{i}=1}{\Sigma }}{\mathrm{x}}_{\mathrm{m}+\mathrm{i},\mathrm{n}+\mathrm{j}}wij+b)(0\le \mathrm{m}\le \mathrm{M},0\le \mathrm{n}\le \mathrm{N}) (1)$$

$$\text{ Feature map Size }=(\mathrm{N}-\mathrm{F}+2\mathrm{P})/\mathrm{S}+1 (2)$$

2) Max Pooling Layer

After convolutional layers, CNNs frequently employ the pooling layer operation (Sowmya et al., 2023). The goal is to reduce the dimension, also known as down sampling. For the pooling layer, max pooling is used which takes the maximum values from the feature map. Max Pooling is shown in equation (3) (Sowmya et al., 2023), in Figure 4.

$${\mathrm{f}}_{\mathrm{pool}}=\mathrm{Max}({\mathrm{x}}_{\mathrm{m},\mathrm{n}},{\mathrm{x}}_{\mathrm{m}+1,\mathrm{n}},{\mathrm{x}}_{\mathrm{m},\mathrm{n}+1},{\mathrm{x}}_{\mathrm{m}+1,\mathrm{n}+1})(0\le \mathrm{m}\le \mathrm{M},0\le \mathrm{n}\le \mathrm{N}) (3)$$

In the formula, f_pool represents the maximum pooling result of feature graph.

Then, features are fed into a fully connected (FC) layer which uses flattening. Flattening is used to convert all the resultant 2-dDimensional arrays from pooled feature maps into a single long continuous linear vector (Sowmya et al., 2023). Dropout and activation functions are connected to the FC layer. Dropout is used to solve the over fitting problem and improve the generalization error. An activation function decides whether a neuron should be activated or not. It helps to normalize the output of each neuron to a range between 1 and 0. The activation function used in 2D-CNN are the rectified linear unit ReLU, Softmax, tanH and the Sigmoid functions. The ReLU activation function is used in this work. The advantages of ReLU include the fact that it replaces negative values in the feature map by zero, moreover maximum threshold values are infinite, so there is no issue with the vanish gradient problem, the output prediction accuracy and r efficiency are maximum, speed is fast compared to other activation functions. Relu activation function is shown in equation (3) and it is shown in Figure 5:

$$\mathit{Max}(0,x)=x,x<0,=0,\mathrm{x}<0 (4)$$

3) Softmax Layer

Softmax is used in the last layer of a CNN to output a probability distribution over the possible classes. This is because the last layer of a CNN is responsible for making the final prediction about the class of an input image.