Manual detection is crucial for managing human-elephant conflict, especially on roads or railway or human localities. Cloud-based elephant detection using YOLOv7 could mitigate conflict and provide surveillance for elephant transgression. Combining units could address larger issues. To address this issue, two key tasks are proposed: First: Detecting elephants on the railway track or highway or nearby human localities using improved YOLOv7 model, which is integration of Visual Attention Network (VAN) layer with the YOLOv7 to recognize elephants in real-time. Second: Notifying the relevant authorities. This paper examines the efectiveness of object detection methods for identifying elephants, focusing on diferent variants belonging to YOLO (You Only Look Once). After comparing these versions, the improved YOLOv7 model demonstrated superior performance on a custom Elephant Detection (ED) dataset. The model was trained on a combination of free and custom elephant datasets, with cloud-based cameras capturing images from multiple locations. The model attained an impressive validation accuracy of 97%.
Initial investigations indicate most collisions take place in particular ’hotspot’ areas where elephant pathways cross roads or railway tracks. Often, these elephant/vehicle collisions occur because drivers do not have enough time to react at sharp turns, at night driving or under adverse weather conditions. A vision-based detection system was designed and tested in a prototype early warning system to address this problem. Driven by the initial results, detection accuracy is shown to be satisfactory under extremely varying lighting conditions under the assumption of having extensive training datasets that capture many challenging scenarios. The prototype has been shown to be robust and reliable as a whole.
This paper is structured as follows: Section 1 outlines the introduction, Section 2 reviews related works, while Section 3 introduces the proposed system with its components, while Section 4 presents the experimental results, and finally, Section 5 represents conclusion followed by references.
Ecological balance depends on the presence of wild animals in the Earth. Many studies have been
done in this topic yet more studies are still needed. One such problem is the problem of animal vehicle
collisions (AVCs), a serious problem for biodiversity, which Saxena A., et al proposed for an intervention
that uses deep learning techniques for wildlife detection and avoidance of collisions [
Sugumar and Jayaparvathy [
A comparable result on various tasks, including image classification, object detection, semantic
segmentation, panoptic segmentation, pose estimation, etc. achieved by VAN [
Each algorithm has its advantages and disadvantages, and the choice of which algorithm to use depends on the specific requirements of the problem. This paper also explores detection methods based on deep learning that are used to identify elephants on the streets, railway tracks or human localities as well as sending an informational alert to the concern authorities, such that appropriate action might be taken.
The dataset utilized in this research was sourced through the internet for the purpose of elephant detection. Images were collected featuring elephants in diverse orientations, lighting conditions, and backgrounds. Various deep learning techniques can be applied to enhance these images. Additionally, a Python script was employed to extract images from videos. Figure 1 illustrates the process of data collection and annotation for the elephant dataset using the LabelImg tool to prepare for training.
Basically, there are freely available datasets of elephant’s images are present on some open-source free
websites, like: “The Aerial Elephant Dataset" [
Data preprocessing is one step to improve the quality of data. However, this process involves organizing plus processing raw data to generate results that are readable as well as easily available. Complexity, accuracy and suficiency are common challenges to image data. However, the importance of the image data processing is something that has been under investigated in data science. This includes grayscale conversion, normalization, data augmentation and image standardization. In this study, the data augmentation is employed to expand as well as adjusted the dataset size plus resized the images accordingly.
Towards preparing the training dataset, the frames are extracted from the videos. However, most of these frames did not contain any elephants. To ensure robust results, the frames that only retained where confirmed elephants included, discarding the others.
To annotate the dataset, the “LabelImg" open-source tool is used. Users draw bounding boxes around areas of interest and label the classes. After saving, a text file is generated, where the first decimal value indicates the class ID is listed first, followed by the x-axis and y-axis centers, width as well as height. The dataset is subsequently split into training as well as testing portions for additional processing. LabelImg can export annotations in various formats, including YOLO. If the dataset is labeled with LabelImg, annotations can be directly exported in YOLO format, simplifying the conversion process.
LabelImg provides key features for eficient image annotation: • User-Friendly Interface: Built with Python and Qt. • Annotation Formats: Supports PASCAL VOC XML, YOLO, and CreateML. • Annotation Modes: Detection, segmentation, and classification. • Pre-defined Classes: Easily create and manage classes for labeling. • Keyboard Shortcuts: Quick actions for bounding boxes, saving, and navigation.
• Remote Access: Annotate images directly from remote servers.
The vast amount of data collected from Internet of Things (IoT) devices must be stored on a secure server, with cloud computing playing a crucial role in this process. Once the data is processed and analyzed, it helps identify electrical faults, errors, and other system issues more efectively.
An internet connection is crucial for communication, with each physical object assigned an Internet Protocol (IP) address. However, as device usage grows, the limited number of available IP addresses will become inadequate, prompting to explore alternative identification methods.
Deep learning is a subset of machine learning within artificial intelligence that utilizes neural networks to learn from unstructured or unlabeled data. Also known as deep neural learning or deep neural networks, it automates learning processes and is inspired by the structure of the human brain.
YOLOv7, built on the Extended Eficient Layer Aggregation Network (E-ELAN), enhances speed and accuracy with an optimized layer design. Key features include: • Model Scaling: Customizes depth, width, and resolution for various tasks. • Auxiliary Head: Provides extra supervision during training. • Enhanced Loss Function: Boosts training eficiency. • High Performance: Achieves faster, more accurate inference than YOLOv5 and YOLOv4, with fewer parameters and lower computational costs.
This work proposes a framework designed for real-time processing elephant identification plus alert creation to protect both humans and elephants. Cloud services are utilized to connect cameras deployed in various hotspots, with captured images kept in a database for model inspection. For actual deployment of this trained model, a local cloud is utilized to host the model and establish the connection with the area’s cameras, ensuring protection against elephant attacks. This local environment is then incorporated into a global cloud via internet, forming an indispensable component of the system.
The model consists of two phases. In the first phase, the YOLOv7 model is optimized with adjustable metrics using samples from an open-source database, incorporating a Visual Attention Network (VAN) layer. The developed models are evaluated using data from AI-connected cameras positioned near localized hotspots, generating diferent datasets from various locations. Performance variations from the optimized models are analyzed. During the second phase, the trained model is implemented alongside a local cloud server to detect elephants in the area. Upon detection, alerts are sent to local forest authorities and the community as a warning. This Proposed Model states that an integration of YOLOv7 with Visual Attention Network (VAN) will yield an improved version of YOLOv7.
In real time also the model will work eficiently with real time data captured through webcam. This deployment is depicted in Figure 2.
The experimental results compile various studies conducted over the past years in animal reidentification and attribute prediction, unifying them under a common framework. The recognition systems developed for diferent animal species share a similar algorithmic design approach. While the original publications in Machine vision focused on technical aspects belonging to these algorithms, this work examines them from an application perspective. It contextualizes these studies in relation to one another, emphasizing their shared approach and their capacity to incorporate lifelong learning techniques that can enhance the decision models used.
New results are presented for identifying elephants in camera trap videos, showcasing advanced capabilities for monitoring animals beyond individual identification. These advancements include predicting individual attributes and implementing lifelong learning with human is the approach. The latter is increasingly important for real-world applications and long-term monitoring.
To train a model on Google Colab, start by creating fresh computer notebook, setting change runtime type to GPU, and running code to clone this “YOLOv7 repository" plus install necessary segments. Download and extract the dataset, then acquire the “YOLOv7" pretrained weights (yolov7.pt) for faster training. Ensure paths are correct inside ‘data.yaml’ record within ‘yolov7/dataset’ directory. Once training completes, optimal weights are stored within the ‘runs’ directory. With this weight file ready, you can proceed with evaluation and inference using the ‘detect.py’ script to identify elephants in images.
To train the YOLOv7 model, utilize Google Colab by creating fresh computer notebook, setting change runtime to GPU, and running this provided code to clone this YOLOv7 repository plus install necessary segments. Download and extract the dataset into a folder, then obtain the “YOLOv7" weight ifle, “yolov7.pt," for finetuning pretrained weights instead of starting from scratch. Create a data.yaml ifle to specify class names and verify that paths inside data.yaml record located in yolov7/dataset directory. After training, best weights will be stored inside ’runs’ directory. After obtaining this weight ifle, one can apply “detect.py” script to perform inference plus identify elephants in images.
The attention framework is widely employed in machine learning plus deep learning algorithms. [
Components overview: 1. Input Image: The original image or video frame to be processed for object detection. 2. YOLOv7 Object Detection Module: Detects objects in real-time, outputting bounding boxes and class labels for detected objects. 3. Extract Regions of Interest (ROIs): Crops areas of detected objects based on YOLOv7’s bounding boxes for focused analysis. 4. Visual Attention Network (VAN): Applies attention mechanisms to ROIs to enhance feature extraction, focusing on relevant areas and improving representation. 5. Post-Processing: Combines VAN’s outputs with YOLOv7’s results, refining detection accuracy with attention maps or enhanced bounding boxes. 6. Final Output: An image with detected objects and VAN’s enhancements, such as attention maps or refined bounding boxes.
Conclusion: This integrated approach leverages YOLOv7’s real-time detection and VAN’s advanced attention mechanisms to improve object detection by isolating ROIs and applying targeted feature extraction, enhancing accuracy and performance in computer vision applications.
Various hyperparameters, such as scaled image size and batch size, can significantly influence the identification accuracy of the Improved YOLOv7 model. Metrics such as precision, recall, F1-score, and accuracy are commonly used to evaluate the model’s performance. Precision measures the ratio of correct predictions to the total number of positive predictions (false positives included), while recall assesses the proportion relating to correct predictions to the total actual positives (including false negatives). Accuracy is calculated based on both the count of accurate predictions also the count of error ones. Overall accuracy of the proposed model was computed using the formula provided in Equation (1).
An object detection module was developed to identify elephants in localized hotspot areas. The accuracy and speed of the proposed Improved YOLOv7 model were compared to those of the earlier version of YOLOv7 for elephant detection. Before testing, the YOLOv7 algorithms were trained and validated using an elephant detection dataset. The confusion matrix of elephant detection in improved YOLOv7 and YOLOv7 has been shown in Figure 4 and Figure 5 respectively.
There are mainly two classes i.e., “elephant" and “not an elephant". The total 5,535 images of prepared customed dataset have been taken for evaluating the models, which includes 4,324 elephant images and 1,211 non elephant images.
The models have been trained with 3,165 number of original elephant images and 710 number of non-elephant images. The image dataset details for models evaluation has been shown in Table 1.
The testing or validation of the models have been done with 1,660 images, out of which 1,159 images are positive, which contains elephants and 501 images are negative, which does not contain any elephant portrait. Table 2 depicts the results of the experiments.
The accuracy of improved YOLOv7 during training shows 98%, while during testing or validation shows only 97%. The detailed individual results of the models have been shown in the Table 3 for YOLOv7 model and Table 4 for the improved YOLOv7 model.
Correct predictions means true positives plus true negatives of the model during testing phase. Whereas, total predictions means summation of all predicted values. The object detected image of elephant is shown in Figure 6 evaluated in YOLOv7 model.
The object detected image of elephant evaluated in improved YOLOv7 model is shown in Figure 7.
Accuracy: It is defined as the ratio of correct predictions made by the proposed model to the total number of predictions. This metric is particularly efective when the classes of the target variable are balanced within the dataset. It can be represented as follows.
= =
In this paper an integration of YOLOv7 with Visual Attention Network (VAN) is stated, that yields an improved version of YOLOv7. Incorporating VAN into YOLOv7 enables fast inference backed by the ability to understand what it sees. Combined, accuracy, reliability, and velocity optimize in complex conditions, for various actual-time domains. The deep learning algorithm is combined with a Visual
Attention Network to improve the control of parameters, selection of parameters, and convergence rate to improve the image detection and classification models. In order to prove the eficacy of the proposed method, it was implemented on an open-source elephant dataset. Both the experimental outcome and computational analysis suggest that the enhanced YOLOv7 has robust optimization features and is capable of excelling in elephant recognition and diferentiation notably in the concerns of precise, grasp, and F1 measure. The proposed model achieved training accuracy of 98% and validation (or testing) accuracy of 97%.
The author(s) have not employed any Generative AI tools.