Tiger detection using light and efficient models

• Main Author: Somdath Kshitij Agarwal
• Other Authors: Deepu Rajan
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2023
• Subjects: Engineering::Computer science and engineering::Computing methodologies::Artificial intelligence
• Online Access: https://hdl.handle.net/10356/166078
• Institution: Nanyang Technological University

In recent years, computer vision has emerged as an important tool for wildlife conservation. The ability to automatically detect and track animals in their habitats helps gathering crucial data for their preservation efforts, such as population monitoring, habitat assessment and deployment of anti-poaching measures. However, computer vision deployment for wildlife conservation has its own set of challenges to overcome, especially when it comes to the limited resources available in remote and inaccessible areas. There is a strong need for lightweight object detection models which are fast and efficient and can work under limited computational resources and power. Edge cameras are low-power, low-cost devices that can help capturing images and videos locally, without the need for a continuous internet connection. This makes them ideal for use in wildlife conservation. They typically have limited computational resources, which means that object detection models need to be small and efficient to work on these devices. However, the detection models with highest accuracy are generally based on either a two-stage approach such as Fast R-CNN [12], or one-stage approaches such as RetinaNet [3] or models with backbones which are deep and complex, such as SSD [4] which make them difficult to adopt on edge devices. In this project, we propose a new Feature Pyramid Network (FPN) [7] based architecture with a ShuffleNetv2 [9] backbone for efficient and quick tiger detection task. We have chosen Amur Tiger detection as our use case as these tigers are an endangered species, whose conservation is vital for maintaining ecological balance of their habitats. Our FPN based model uses pyramid of feature maps with different scales to detect objects at different resolutions. The ShuffleNetv2 backbone is exploited to extract features from the input images, which are then passed through FPN layers for object detection. This model is trained on the publicly available ATRW dataset [5] and achieves superior performance in terms of accuracy and speed. The main advantage of our proposed architecture is its low computational cost. The ShuffleNetv2 is a highly efficient architecture that has shown to achieve good performance on a variety of tasks while using significantly fewer parameters than other state-of-the-art models [9], which makes it a promising choice for edge devices. Our model makes a promising choice for fast and efficient tiger detection in the wild while being deployed on the edge-devices in real-time. Code and model available at : https://github.com/kshitij0807/wild-tiger-detection