As for all birds, the behavior of chickens is largely determined by environmental conditions. In many production systems, light intensity is low and red feather strains have low contrast with the background, making it impossible to use conventional image segmentation techniques. On the other hand, studies of chicken behavior, even when using video camera resources, depend on human vision to extract the information of interest; and in this case, reduced samples are observed, due to the high cost of time and energy. Our work combined the use of advanced object detection techniques using YOLO v4 architecture to locate chickens in low-quality videos, and we automatically extracted information on the location of birds in more than 648 h of footage. We develop an automated system that allows the chickens to transition among three environments with different illuminations equipped with video cameras to monitor the presence of birds in each compartment, and we automatically count the number of birds in each compartment and determine their preference. Our chicken detection algorithm shows a mean average precision of 99.9%, and a manual inspection of the results showed an accuracy of 98.8%. Behavioral analysis results based on bird unrest index and permanence time indicate that chickens tend to prefer white light and disfavor green light, except in the presence of heat stress when no clear preference can be observed. This study demonstrates the potential of using computer vision techniques with low-resolution, low-cost cameras to monitor chickens in low-light conditions.
Keywords: YOLO; environmental stress; precision poultry farming; welfare.