利用深度卷積神經網路於商業雞舍進行雞隻活動力偵測

Automatic detecting the activity levels of chickens in commercial chicken farm using deep convolutional neural networks

粘書耀(Shu-Yao Nien)

指導教授 ： 郭彥甫

國立臺灣大學/生物資源暨農學院/生物機電工程學系/碩士(2020年)

https://doi.org/10.6342/NTU202003789

摘要

雞肉是國人飲食中的重要蛋白質攝取來源。在2018年，雞肉產品占臺灣畜牧業總銷售產值的26.32%，是國內相當重要的農業經濟來源。然而雞隻容易因為熱緊迫、打架與生病等原因而減少活動量，進而出現活動力低下的雞隻，因此早期偵測雞隻的活動力對於家禽養殖相當重要。在傳統的管理方式下，飼主以人工巡邏雞舍是每日必須進行的工作。但對於雞隻數量龐大的商業雞舍而言，以傳統管理方式相當耗時、勞力密集且仰賴飼主經驗，此外管理上的失誤可能造成疾病發生與利益損失。故本研究採取使用嵌入式影像蒐集系統、深度學習模型與物件追蹤演算法進行商業雞舍中的雞隻活動力監測。在此研究中，利用自製的嵌入式系統以每秒五幀的頻率來獲取雞舍內的雞隻影片，並透過更快速區域生成卷積神經網路(Faster R-CNN)進行影片中每幀影像的雞隻偵測及定位，接著使用空間分布指數(SDI)量化雞隻影像中的空間分布情形，最後利用SORT演算法與訓練模型所生成的邊界框進行雞隻的軌跡預測及活動力判斷。研究中的Faster R-CNN模型在雞隻偵測上達到90.96%的平均精度(AP)，此外SDI的表現可預估影像中雞隻的分布情形。另外SORT演算法在多物件追蹤準確度(MOTA)、多物件追蹤精確度(MOTP)與識別F1值(IDF1)則分別可達到92.1%、84.4%與89.5%。而雞隻軌跡與運動量的分析結果則可以做為尋找活動力低下雞隻的參考。

關鍵字

雞隻活動力 ； 雞隻追蹤 ； 卷積神經網路 ； 深度學習 ； 嵌入式系統 ； 物件偵測 ； 空間分布

並列摘要

Chicken is a major source of dietary protein. In 2018, chicken production accounted for 26.32% of total animal husbandry sales in Taiwan. Chickens may reduce their movements and become inactive when they suffered from heat stress, injured in fights, or infected by diseases. Thus, early detecting the activity levels of chickens is essential to chicken farming. Conventionally, chicken farm patrol is a routine for chicken farm owners to monitor the activity levels of chickens. A typical chicken farm usually contains thousands of chickens, making patrol laborious and time-consuming. Also, naked-eye observation may be prone to error due to fatigue. This work proposed to detect the activity levels of chickens in a commercial chicken farm using embedded systems, deep learning, and tracking algorithm. In this study, embedded systems were designed to acquire the videos of the chicken farm. A faster region-based convolutional neural network (Faster R-CNN) was developed to detect and localize the chickens in the frames of videos. The spatial dispersion of chickens in images were quantified using spatial distribution index (SDI). Simple online and realtime tracking (SORT) was subsequently used to track chickens and to identify the activity levels of chickens using the locations provided by the CNN model. The trained Faster R-CNN model reached an AP of 90.96% in chicken detection. The performance of SDI could evaluate the distribution of chickens in images. SORT reached an MOTA of 92.1%, an MOTP of 84.4%, and an IDF1 of 89.5% for ID metrics. The trajectories and movements of chicken could be the standard for finding inactive chickens.

並列關鍵字

Activity levels of chicken ； Chicken tracking ； Convolutional neural network ； Deep learning ； Embedded system ； Object detection ； Spatial dispersion

參考文獻

Bewley, A., Ge, Z., Ott, L., Ramos, F., Upcroft, B. (2016, September). Simple online and realtime tracking. In 2016 IEEE International Conference on Image Processing (ICIP) (pp. 3464-3468). IEEE.

Caubet, Y., Richard, F. J. (2015). NEIGHBOUR-IN: Image processing software for spatial analysis of animal grouping. ZooKeys, (515), 173.

Cheng, D., Rong, T., Cao, G. (2019, August). Density Map Estimation for Crowded Chicken. In International Conference on Image and Graphics (pp. 432-441). Springer, Cham.

Colles, F. M., Cain, R. J., Nickson, T., Smith, A. L., Roberts, S. J., Maiden, M. C., Lunn, D., Dawkins, M. S. (2016). Monitoring chicken flock behaviour provides early warning of infection by human pathogen Campylobacter. Proceedings of the Royal Society B: Biological Sciences, 283(1822), 20152323.

Council of Agriculture, Executive Yuan, Taiwan. (2019). Agricultural Statistics Yearbook 2018. Taipei, Taiwan: Council of Agriculture, Executive Yuan.