Analyzing chicken activity level under heat stress condition using deep convolutional neural networks

Abstract

<b><sc>Abstract.</sc></b> Chicken is a major source of protein. The production value of chicken accounted for 39.65 % of total animal husbandry sales in Taiwan. Because Taiwan is a subtropical country, heat stress may occur on chickens due to the high temperature and humidity in summer. Without appropriate care, heat stress may cause sudden death of chickens. Chickens may become inactive when they suffer from heat stress. Therefore, detecting the activity levels of chickens is a straightforward way to measure their risk of heat stress. This study aimed to measure the activity levels of chickens under various temperature conditions using computer vision. In the experiment, chicken eight weeks of age were raised in an experiment chicken house. The chickens were treated with normal or high temperature conditions. A raspberry-pi and a web camera were used to acquire overhead video of the chickens at a rate of five frames per second. A You Only Look Once—version 4 tiny model (YOLOv4-tiny) was trained to detect and to localize chickens in the video frames. Simple online and real-time tracking (SORT) was used for chicken tracking and quantify the movement from the detected location provided by YOLOv4-tiny model. The mean average precision of trained YOLOv4-tiny model achieved 93.20% in chicken detection. SORT achieved a multiple object accuracy of 99.2%, a multiple object precision of 84.3%, and an identification F1-score of 96.6%. The activity level found in different temperature conditions can be quantified as chicken‘s risk of heat stress.