Real-Time Detection of Endangered Birds Based on Lightweight Networks and Zero-Copy Kernel Processing Unit Acceleration

Intelligent monitoring of endangered birds is a crucial component of ecological conservation. However, traditional deep learning birds detection methods on edge terminals or offline devices are constrained by computational resources, memory, and energy consumption, making it difficult to sustain long-term real-time monitoring in outdoor environments. This study presents a real-time endangered birds detection edge system for endangered birds based on a lightweight network and a zero-copy mechanism. Firstly, we designed and implemented an effective lightweight detection model called Vanilla-BirdNet. Constructed upon the YOLOv12 framework, this model incorporates a lightweight VanillaBlock detection head, achieving a computational efficiency of 5.4 GFLOPs with high detection accuracy (0.924 mAP@0.5:0.95). We utilized the Shape-IoU loss function to optimize the model, effectively accelerating convergence and improving recall to 0.945, thereby significantly enhancing localization precision for birds with varied postures. Secondly, at the system level, we introduced a zero-copy data mapping Kernel Processing Unit (KPU) inference mechanism. Leveraging the hardware characteristics of the Artificial Intelligence (AI) System-on-Chip (SoC) K230 chip, this mechanism effectively reduces memory copying and bandwidth consumption, leading to lower end-to-end inference latency. Finally, the entire system was successfully deployed on the K230 edge computing platform. Validation results on a real-world dataset demonstrate that the edge platform achieves real-time detection performance of 14.2 frames per second (FPS) (with 320 \(\times \) 320 input) at a power consumption of less than 2.4W. This represents an approximately 29% improvement compared to baseline models, showcasing its immense potential for long-term automated deployment in wild environments.