Development of an efficient framework based on Deep Learning for Wildlife Surveillance using Robot

Abstract

Wildlife mortality on train tracks represents a significant conservation challenge, as expanding rail networks increasingly intersect with vital animal habitats, leading to frequent and often devastating collisions. Conventional methods for gauging this impact, such as on-the-ground counts, are often limited in their reach and are susceptible to inaccuracies caused by natural decay and predator activity. In contrast, contemporary wildlife tracking methods offer a far more robust solution. By employing remote sensing, camera networks, and audio recording devices, it is possible to achieve continuous and extensive data collection across wide areas. This reduces the need for constant human presence, minimizing disruption to vulnerable species. The capability to collect real-time data allows for the swift identification of high-risk zones and the evaluation of preventative measures, leading to more data-driven and timely conservation efforts. Deep learning algorithms are rapidly becoming crucial for interpreting wildlife tracking data, significantly enhancing the understanding of animal habits and mortality trends. By training neural networks on diverse datasets, including visual and auditory records, recognition and categorization of animals can be automated, even under difficult circumstances. This allows for the efficient processing of large datasets, revealing subtle patterns that would otherwise go unnoticed. By training image classification models on carefully labelled images, one can identify species, track movement, and analyse behaviour, providing essential information for understanding the factors that contribute to train-related animal deaths. Furthermore, it enables the assessment of the effectiveness of mitigation strategies, such as animal overpasses and deterrent systems, by analysing how animals interact with these structures. Autonomous mobile robots equipped with high-resolution cameras and onboard processing capabilities can traverse these areas, capturing vast amounts of visual data. These robots, guided by pre-trained image classification models, can autonomously detect and categorize animals, providing real-time insights into population dynamics and mortality risks. The principal aim of this research is to construct a robust, deep learningdriven framework for automated wildlife surveillance using robotic platforms. This investigation centres on creating a system that enables autonomous robots to navigate and analyze environments, capturing and processing visual data to identify, categorize, vi and track animal. This proposed approach allows for continuous, non-invasive monitoring, reducing human presence and minimizing disturbance to sensitive species. In the first proposed approach, the autonomous robot will sense the animal and capture the image using installed camera. This image is used for image segmentation and classification. To achieve this, Mask-RCNN for feature extraction and prediction with the concept of AMP is used. Unlike existing schemes, the proposed scheme is capable enough for fast computation and maintains accuracy that will be efficiently implemented in a real-time scenario such as wildlife detection. The proposed model is implemented on ROS based mobile robot with Raspberry Pi4. The results obtained using the proposed scheme obtained a mean Average Precision value of 85.47% and an F1 score of 87.73% with a precision value range between 92% to 99%. The second proposed approach introduces PAW: Prediction of wildlife animals using a robot under adverse weather conditions. In this, implementation of image dehazing is performed to extract features from images captured under adverse weather conditions. This approach is an improvement on previous approach, in which image dehazing algorithms are incorporated. In this approach, two main improvements have been done – one improvement is an addition of images in the dataset and other is dehazing before extracting features and classifying the image. In this, dataset of images clicked in bad weather conditions is used, such as mist, haze, smog, and fog, often suffer from poor visibility. To train and test the model, synthetic images consisting of haze and fog were generated using GIMP tool. The results obtained using the proposed scheme obtained a mean Average Precision value of 88% and an F1 score of 92%. In the third proposed approach: Improved Packet Delivery and Energy Consumption centric RSA (IPDEC-RSA), the traditional reptile search algorithm is modified to ensure fast and reliable message delivery with reduced energy consumption and minimum routing distance. The system operates in four phases: first phase is to deploy static robots with sensor nodes in clusters, each with a Cluster Head (CH). In the second phase, optimal CHs are selected for routing based on RSA approach. In third phase, messages are routed through the selected CHs to the Base Station (BS). In fourth stage, data is collected at the BS and transmitted to a cloud server or SMS using MQTT. The model is tested and validated on synthetic hazy dataset, which shows the effective results in controlled real-time implementation.