基于改进YOLO v5-StrongSORT的屠宰场猪只精准计数方法

Accurate Counting of Pigs in Slaughterhouses Based on Improved YOLO v5-StrongSORT

摘要:

猪只计数是屠宰场生产管理、资产估计的重要环节。针对当前屠宰场猪只数量统计过程中人工计数耗时长、错误率高的问题，提出了一种基于改进YOLO v5-StrongSORT的屠宰场猪只精准计数方法。首先，在改进YOLO v5模型中引入真实宽高损失与纵横比以提升损失函数性能，并在Neck层引入高效通道注意力机制(Efficient channel attention, ECA)，提升模型在复杂环境下的识别能力。然后，基于StrongSORT构建检测机制实现对猪只的重识别。最后，基于StrongSORT提出了一种区域ID信息检测的猪只计数方法。试验结果表明，改进YOLO v5模型对猪只识别精确率为93.78%，召回率为91.98%，平均精度均值为96.29%，识别速度为500f/s，较YOLO v5s模型召回率提高1.14个百分点，平均精度均值提高0.89个百分点，识别速度提高85.0%。将改进YOLO v5与StrongSORT区域计数方法结合进行猪只计数的准确率为98.46%，计数速率为73f/s，较人工计数准确率提高1.54个百分点，较原始模型计数准确率提高9.23个百分点，计数速率提高21.87%。本研究猪只计数方法具有较高的计数精度，适用于屠宰场猪只数量统计。

关键词:猪只识别;区域计数;改进YOLO v5;损失函数;注意力机制;StrongSORT

Abstract:

Pig counting plays a crucial role in the management of slaughterhouse production and the estimation of assets. In response to the existing challenges of labor-intensive manual counting and elevated error rates within the pig counting processes of slaughterhouses, a meticulous pig counting methodology was introduced, leveraging an improved integration of YOLO v5 and StrongSORT. Initially, the improved YOLO v5 model incorporated real aspect loss and aspect ratio to enhance the performance of the loss function. Additionally, an efficient channel attention (ECA) mechanism was introduced into the Neck layer to augment the model’s recognition capabilities in complex environments. Subsequently, a detection mechanism was constructed based on StrongSORT to facilitate the re-identification of pigs. Finally, a pig counting method utilizing area ID information detection was introduced based on the StrongSORT framework. Experimental results demonstrated that the enhanced YOLO v5 algorithm achieved a pig recognition accuracy of 93.78%, a recall rate of 91.98%, and a mean average precision (mAP) of 96.29%, with a recognition speed of 500 frames per second (f/s). This represented a significant improvement of 1.14 percentage points in recall, 0.89 percentage points in mAP, and an 85.0% increase in frame rate compared with that of the YOLO v5s model. The accuracy of combining the improved YOLO v5 with the StrongSORT area counting method for pig counting was 98.46%, and the counting speed was 73f/s, which was 1.54 percentage points higher than the manual counting accuracy, 9.23 percentage points higher than the original model counting accuracy, and 21.87% higher than the counting speed. The pig counting method proposed demonstrated high accuracy and was well-suited for the enumeration of pigs in slaughterhouse settings.

Key words:pig identification;region counting;improved YOLO v5;loss function;attention mechanism;strongSORT