Development of method for lameness detection based on depth image analysis

To maintain the physical condition of sows frequent observations are necessary. Lameness is a common concern in group-housed systems. However, current methods are done manually by subjective methods, thus are difficult to complete. One alternative would be to automate the process by analyzing images generated by depth cameras. The present work aimed to develop and test a method for early detection of lameness in sows, adapting the kinematics method using top down depth cameras without the aid of reflective markers. Depth images were processed by dividing the animals into five body regions: head, left and right shoulders, and left and right hips. The centroid of each position was calculated, and their heights were recorded. Average, maximum, and minimum height at each of the five regions was calculated. The animal's velocity was also measured by calculating the Euclidian distance between positions of the animals' body centroid (not considering its head) over the time between frames. The curves of height by time obtained for the centroids of all five regions were plotted and analyzed. Time and length for each step was also computed. Preliminary results indicate lameness level was best correlated with number, time, and length of steps for each of four regions (left and right shoulders and left and right hips); total walk time; and number of local maxima for the head region. With the automation of lameness detection, it could be possible to have better insights on the physical condition of sows and aid on better and faster management decisions.