# Enhancing Wildlife Monitoring: An Advanced AI Approach for Accurate Giant Panda Behavior Detection and Conservation Insights

## TL;DR

This paper introduces an AI model called PandaSlowFast that improves the accuracy of monitoring giant panda behaviors in the wild, with potential applications for other endangered species.

## Contribution

The novel PandaSlowFast AI model with channel attention and optimized convolutions achieves higher accuracy and practical edge deployment for wildlife monitoring.

## Key findings

- PandaSlowFast achieved 85.38% mean average precision (mAP) on a new dataset of wild panda videos.
- A compact version of the model maintained 85.16% mAP and ran at 3.2 frames per second on a Raspberry Pi 4.
- The model's design includes channel attention, small-kernel convolutions, and the Adaptive SwisH activation function.

## Abstract

Protecting endangered species like giant pandas requires constant monitoring, but current methods often struggle with complex forest environments. This study introduces PandaSlowFast, an improved AI model that automatically recognizes wild panda behaviors from video footage. By enhancing how the model processes motion patterns and fine image details, we achieved 85.38% accuracy on a new dataset of long-term monitoring videos—significantly outperforming existing approaches. A compact version also ran effectively on a simple edge device like Raspberry Pi, making it practical for on-site deployment. This technology supports real-time wildlife monitoring and can be adapted for other rare species.

As global demands for nature reserve management intensify, intelligent monitoring has become a pivotal trend. Integrating artificial intelligence with infrared camera traps enables automated analysis of endangered species behavior, providing timely insights for conservation. However, complex habitats often degrade the performance of existing detection technologies. Focusing on the giant panda—a flagship conservation species—we constructed a novel dataset from long-term field monitoring videos and developed an improved PandaSlowFast network. Our model employs channel attention to enhance temporal features, uses small-kernel depth-wise convolutions and dilated convolutions to expand receptive fields for spatial feature extraction, and introduces the Adaptive SwisH activation function to improve adaptability and training stability. The results show that PandaSlowFast achieves 85.38% mean average precision (mAP), outperforming existing methods. An FP16-quantized version maintains comparable accuracy (85.16% mAP) while running at 3.2 frames per second on a Raspberry Pi 4, demonstrating practical deployability for on-site monitoring. This work provides technical support for intelligent panda behavior analysis and offers a transferable methodology for monitoring other rare species, contributing to biodiversity conservation.