NAM-YOLOV7: An Improved YOLOv7 Based on Attention Model for Animal Death Detection

Traffic on the roads is becoming more congested as the population grows. When the animals try to cross the roads on a traffic road it results in animal vehicle collisions. As a result, animals are decreasing, which lead to animal accidents. Human deaths and animal deaths have both increased dramatically in recent years.

Animals are especially vulnerable to natural disasters such as earthquakes, floods, thunderstorms, heavy rainfall, and fires, which can all be disastrous. Many animals are killed when they are covered in sand and dust as a result of floods in water. Some animal species are killed as a result of fallen trees, heavy wind. Many of the animals were killed as a result of the toxic gases emitted by the factories. Animal deaths are increasing dramatically due to a variety of factors. Recently in Gujarat rains, one person and 90 animals were killed in Anand flooding due to heavy rainfall on July 3 2022. Over 250 cows die due to heavy rains in Odisha in December 2018. Several animals are killed in a gas leak at the LG Polymers factory in Vizag in May 2020.

Dead animals on the road should be removed quickly as early as possible. A dead animal may release bacteria that were previously contained within its body as it decomposes, potentially exposing people to disease-causing pathogens. Dead animals pose a risk to public health, because of offensive odours cause respiratory problems in humans and also other animals. When an animal dies, it is the owner's responsibility to dispose of the carcass in a manner that is not harmful to the environment or public health. This is true for farm owners, pet owners, service animal organisations (such as police departments), animal shelters, and laboratories. When people see roadkill, they call their local police or animal control department, and the municipality picks it up. The National Highway Administration and local village secretariats should be held accountable for the removal of dead animals from the roads. It takes some time for the municipality to remove the dead animals. So, to reduce the time delay we can accurately detect dead animals using deep learning and computer vision techniques.

Using computer vision, computers and machines can be taught to understand human action, behaviour, and language in much the same way as humans do. ML includes computer vision, which aims to make sense of images and videos by programming a computer to interpret the visual information contained within the digital data. In order to solve complex problems and make better business decisions, we need to translate data into meaningful insights based on contextual information provided by humans. Deep learning is applied in various areas such as image and video processing, image captioning [1], object-detection [2], natural language processing [3] etc. In DL, artificial neural networks are used to mimic the functioning of the human brain. As the human brain does, machines can learn through deep learning. Deep learning is applied in a wide range of applications in the recent times. Object-detection is the process of “locating” and "classifying" Objects. A procedure designed to detect objects consists of locating relevant elements, drawing rectangular bounding boxes around them, and determining each object's class. Object-detection can be done through ML and DL approaches. ML methods include "SIFT", "HOG" and Viola-Jones object detection and these methods are not efficient for accurate detection.

Several deep learning-based object--detection techniques are available i.e., Faster R-CNN, YOLO, YOLO v2, YOLOv3, YOLOv4, YOLOv5, YOLOv6, YOLOv7 and single shot detection (SSD). Each algorithm has some merits and some demerits and which algorithm should be used is determined based on the problem requirements. In this paper also, we used deep learning-based detection methods to detect dead animals on the roads and sending an alert information to the municipal authorities, so that municipality could take appropriate action.

Camera-based techniques have been used to detect animals on roads by many researchers, but they have several limitations as well. Despite the challenges, detection in coloured images has been successfully implemented using cutting-edge deep learning techniques [4-6]. Detection of animals on the road is an important problem that need to be done within some specific time boundary. As a result, the solution must be time efficient.

For dead animal detection we need images. So, we gathered images from the internet and some images are captured through drones. Practically we can use UAVs to detect dead animals. UAVs are used for surveillance, traffic monitoring, and surveys, among other things. In recent years, they have primarily replaced choppers due to their greater mobility and flexibility [7]. UAVs are being used in modern society with real-time deep learning algorithms that are more efficient and accurate. Aerial sensing research using deep neural networks has become increasingly popular in urban, environmental, and agricultural contexts [8].

A Brief analysis on animal detection is displayed in Table 1. The use of an automatic animal detection and warning system can help reduce collisions between vehicles and animals on roads and highways [9-13]. Trampling by animals such as "cows", "monkeys”, “rats”, “deer", "peacocks”, “elephants" and others causes significant crop damage. During the day, crops are easier to protect. At night, crops are more difficult to protect [14, 15]. van Gemert et al. [16] uses animal detection for animal counting. El Abbadi et al. [17], Tan et al. [18], Ulhaq et al. [19] detect the animals and classify the animals.

Table 1. Analysis on animal detection

Van Gemert et al. [16] used dataset containing of 6 videos recorded by “Ascending Technologies Pelican" (quadcopter)with a mounted GoPro HERO 3: Black Edition action camera. El Abbadi et al. [17] developed their own dataset consisting of 12000 images for their study in animal detection. Tan et al. [18] used “(Northeast Tiger and Leopard National Park wildlife dataset)” for animal detection.

In this paper, we claim the following contributions:

·The system takes the video or images as input and continuously checks for the Dead animals.

·Train the model using different algorithms to identify dead animals.

·On a large scale, we validate our process on DAD (Dead Animal Detection dataset) and inform the concerned municipal authorities for dead animals removal.

It is possible to detect objects using either an anchor-based or an anchor--free method in the era of deep learning. The anchor--based detection method comprises a one-stage algorithm and a two--stage algorithm [20, 21]. The one--stage detector generates a "probability of a class" and a "position coordinate value" for the object, while the two--stage detector generates a "region proposal" and a "target boundary" based on the region proposal [22].

Increased network speed and strength in YOLOv7 allow for more efficient label assignment and model training, as well as improved accuracy in object identification and a more reliable loss function. So, in this study, we used YOLOv7, to identify, localise, and identify dead animals on roads, highways, and forests.

3.1 YOLOv7 architecture overview

Regarding object detection models, there are many different object detection models that perform well for specific use cases, but the recent release of "YOLOv7", was good enough in both speed and accuracy among all object detectors. In terms of speed and accuracy, the proposed YOLOv7 version E6 outperformed transformer-based detectors like "SWINL Cascade-Mask R-CNNR-CNN". The YOLOv7 outperformed the “Scaled-YOLOv4", "DINO-5scale-R50", "YOLOv5", "DETR", "Deformable DETR", "YOLOX", "YOLOR”, and "Vit-Adapter-B". The main key points of YOLOv7 architecture includes

Extended Efficient Layer Aggregation Network: E-ELAN focuses primarily on the computational density and model architecture parameters. The main advantage of ELAN was that it allowed a deeper network to learn and converge more effectively by controlling the gradient path.

Model Scaling for Concatenation based Models: For concatenation-based models, the model scaling method involves scaling depth in a computational block and width scaling in the remaining transmission layers.

Planned re-parameterized convolution: "Rep-Conv” can be replaced by a RepConvN layer with no identity connections.

Coarse for auxiliary and fine for lead loss: In this label assigner, predictions for heads and ground truth labels are used together to obtain labels for both training heads and auxiliary heads.

3.2 Improved YOLOv7 with normalization-based attention module (NAM)

The attention mechanism is commonly employed in machine learning and deep learning algorithms [23]. The computer vision attention mechanism finds correlations between original data and highlights significant aspects, such as channel attention, pixel attention, multi-order attention, etc. (Figure 6).

NAM is a variation of the image classification CBAM module [24-26]. Figure 7 shows how channel and spatial attention submodules are redesigned. Attention can be rebalanced by modifying the training weights of variance measurements across channels and spatial dimensions. Batch normalization is a scaling factor (BN).

$\mathrm{B}_{\mathrm{OUT}}=\mathrm{BN}\left(\mathrm{B}_{\mathrm{IN}}\right)=\gamma \frac{B_{I N}-\mu_b}{\sqrt{\sigma_b^2+\varepsilon}}+\beta$         (1)

where, $\gamma, \beta$ represent the trainable scale and shift parameters, respectively. $\mu_b$ and $\sigma_b$ are the mean and standard deviation in each batch $b$, respectively. $\gamma$ is the variance of $\mathrm{BN}$.

$\mathrm{M}_{\mathrm{c}}=\sum\left(\mathrm{W}_\gamma\left(\mathrm{BN}\left(\mathrm{F}_{\mathrm{c}}\right)\right)\right)$           (2)

Similarly, its spatial attention module applies BN to pixels in the spatial dimension. It is so-called pixel normalization (PN). It focuses on informational pixels based on scaling factor λ and modifies weights W_λ. The scaling factor value range between 0 and 1. Let us consider one more function F_se R^H×W×NC is the input feature map and M_s represents the output of the spatial attention module. The Output M_s can be expressed as:

$\mathrm{M}_{\mathrm{s}}=\Sigma\left(\mathrm{W}_\lambda\left(\mathrm{PN}\left(\mathrm{F}_{\mathrm{s}}\right)\right)\right)$         (3)

NAM regularizes the loss function to suppress less important weights. The loss can be expressed as:

loss $=\sum_{(x, y)} l(f(x, W), y)+p \sum g(\gamma)+p \sum g(\lambda)$           (4)

where, l(.) and g(.) represent the loss function and l₁ represents norm penalty function, x and y are the input and output, respectively; W is the network weight; and p is the equilibrium penalty factor.

6.png

Figure 6. Improved YOLOv7 with normalization-based attention module

7.png

Figure 7. Architecture of NAM with two attention submodules

3.3 YOLOv7 implementation details

“YOLOv7" model consists of the following training steps.

·For training, we use Google Colab. Create a new notebook, change the runtime type to GPU, and run the code below. This will clone the "yolov7 repository" and install the modules required for training in the colab environment.

·The dataset will be downloaded and extracted into a folder. Once the dataset has been downloaded, we must obtain the "YOLOv7" weight file. Because training from scratch takes longer, we will finetune previously pretrained weights on our dataset. "Yolov7.pt" will be downloaded in our project.

·Create a data. Yaml file with classes and double-check the paths in the ‘data. Yaml’ file inside the ‘yolov7/dataset' folder, as well as the paths in the training data.

·When the training is finished, the best weight is saved at the specified location, which is usually inside the 'runs' folder. Once we have the weight file, we can begin evaluating and inferencing.

·The detect.py script can be used to identify the alive or live animal in the image.

We created an object detection module that detects dead animals on the road. It has not been possible to find an object detection algorithm for dead animals on the road in previous work. For dead animal detection, we compared the accuracy and speed of our proposed YOLOv7 with the earlier versions of YOLO. Prior to testing the YOLO algorithms were trained and validated on the Animal Death detection dataset. According to Table 2, our investigations confirm that both algorithms are good for dead animal detection and for the protection of public health.