Classification of video clips: wildlife detection monitoring system

• Main Author: Gan, Hao Yi
• Other Authors: Lee Bu Sung, Francis
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2024
• Subjects: Computer and Information Science
• Online Access: https://hdl.handle.net/10356/181160
• Institution: Nanyang Technological University

Monitoring wildlife in their natural habitat presents challenges due to the sheer volume of video data captured by motion-sensor cameras, making manual review a time-consuming task. This project presents the development of an automated wildlife-detection pipeline to efficiently process wildlife footage, filtering out non-animal videos and detecting animals in the remaining clips. The core components of this pipeline include an image classification model, referred to as the “filter model”, which extracts videos with animals, and filter videos without, and an object detection model to identify the sighted wildlife. Both models leverage transfer learning and are further trained on pre-processed and augmented datasets to improve accuracy and robustness. Preliminary tests show that the fully automated pipeline that incorporates both models successfully increased efficiency by approximately 180%, streamlining the process of wildlife monitoring, while still ensuring that videos with animals are still accurately located, detecting over 80% of footages with animals. Additionally, a self-training semi-supervised learning process is integrated in the pipeline, where the model creates pseudo-labels using new data that it consumes. This eliminates the need for manual data labelling, limiting human involvement to only checking the correctness of the pseudo-labels. Thus, the pipeline is capable of significantly reducing the effort required for wildlife monitoring.