Developing an automated wild boar and otter detection and monitoring system through video analysis

Abstract

As Singapore continues its urban expansion while maintaining its "City in Nature" vision, human-wildlife interactions have become more frequent, leading to potential conflicts and ecological challenges. This study aims to develop an automated wild boar and otter detection and monitoring system using motion-activated camera traps and artificial intelligence to support stakeholders such as the National Parks Board (NParks) and the Animal Concerns Research and Education Society (ACRES). Through extensive experimentation, YOLOv10b was selected as the optimal object detection model, achieving an impressive mAP@50 of 0.955, F1-score of 0.915, and an inference time of 13.8ms, outperforming DETR and other YOLOv10 variants with a balance of performance and inference speed. Small vision-language models were evaluated for extracting key attributes and generating captions for images with the detected wildlife. Qwen2-VL-2B was chosen over InternVL2.5-2B after experimentation due to its superior attribute extraction and image captioning capabilities. Prompt engineering via system prompts with added context improved F1-scores for 6 out of 8 key attribute extractions and increased the percentage of relevant image captioning responses from 71% to 96%, resulting in more accurate and informative outputs. A simple rule-based approach was used to classify the image’s need to notify stakeholders achieved an impressive F1-score of 0.912. This system serves as a step towards automated wildlife detection and monitoring systems, enabling stakeholders to respond quickly to human-wildlife encounters and early mitigation efforts to reduce risks to both people and wildlife, supporting Singapore’s vision of a “City in Nature”.