兔舍移动巡检机器人轻量化兔只计数方法

Lightweight Rabbit Counting Method for Mobile Inspection Robots in Rabbitries

摘要:

针对集约化笼养兔场中兔只个体密集、遮挡严重以及移动巡检机器人机载计算资源受限的问题,提出一种适用于资源受限条件下的轻量化兔只自动计数方法。 基于 NanoDet-Plus 框架,通过引入改进的 Ghost-Shuffle 单元构建轻量骨干网络、增强型轻量特征聚合网络以及 GeLU 激活函数,结合空间感知计数策略,实现了鲁棒的兔只自动计数。 基于真实商用笼养兔舍环境采集的数据集,在相同实验条件下将所提出模型与多种典型轻量化目标检测模型进行对比实验,并在移动巡检机器人平台上进行部署验证。 实验结果表明,RabbitDet 在参数量只有 7. 3 × 10 5 和浮点计算量为 1. 2 × 10 9 的情况下,实现了 75. 7% 的平均精度均值和 98. 83% 的计数准确率。 与基线模型相比,参数量和计算开销分别下降了 36. 0% 和 31. 0% ,同时计数准确率提升了 0. 53 个百分点。 在移动巡检机器人平台上的推理实验中,与对比的主流模型相比,该方法实现了计数准确度与推理效率之间的更优平衡。 所提出的方法能够在计算资源受限条件下实现复杂遮挡笼养环境中兔只的稳定、精准计数。

Abstract:

Aiming to address the challenges of high rabbit density and severe occlusion in intensive caged rabbit farms, as well as the limited onboard computational resources of mobile inspection robots, a lightweight automatic rabbit counting method suitable for resource-constrained conditions was proposed. Based on the NanoDet-Plus framework, robust automatic rabbit counting was achieved by introducing a lightweight backbone network constructed with improved Ghost-Shuffle units, an enhanced lightweight feature aggregation network, and the GeLU activation function, combined with a spatial-aware counting strategy. Based on a dataset collected in a real commercial caged rabbit house environment, comparative experiments were conducted between the proposed model and multiple typical lightweight object detection models under identical experimental conditions, and deployment verification was performed on a mobile inspection robot platform. Experimental results showed that RabbitDet achieved an mAP of 75. 7% and a counting accuracy of 98. 83% with only 7. 3 × 10 5 parameters and 1. 2 × 10 9 FLOPs. Compared with the baseline model, the parameter size and computational cost were reduced by 36. 0% and 31. 0% , respectively, while the counting accuracy was improved by 0. 53 percentage points. In inference experiments on the mobile inspection robot platform, compared with the baseline models, the proposed method achieved a superior balance between counting accuracy and inference efficiency. The proposed method enabled stable and accurate counting of rabbits in caged environments with complex occlusion under resource-constrained conditions, providing effective technical support for mobile inspection and inventory management in intensive rabbit farms.