Design and validation of a stereo vision system to estimate volume of total mixed rations offered to dairy cattle

Per-animal intake measurements are a valuable component to optimize production, health, and feed efficiency in dairy cattle. The objective of this study was to evaluate depth imaging as a noninvasive monitoring method for estimation of total mixed ration (TMR) amount offered to cows across time. Visual quantification of TMR volume in feed bunks across three unique environmental conditions (Exp. 1) and time (Exp. 2) were used to determine if changes in light conditions and TMR factors affected the accuracy of estimated TMR amount. The study was conducted utilizing an OAK-D Power over Ethernet (PoE) stereo vision camera mounted directly above the tiestalls, which was wire-routed to a computer configured with the Linux Ubuntu operating system. Python and domain-specific open-source software libraries (DepthAI and Open3D) were utilized in tandem with the camera sensor’s output to estimate TMR present in the bunk. Mass, volume, and initial density of the TMR were recorded as manual measurement data to be compared to estimates produced by the camera system. In Exp. 1, varying light exposure, shape of offered TMR, and diet type were evaluated. Diet type and light exposure did not significantly impact system volume estimates; however, shape of TMR did, as diets that were more spread out were perceived to be of less volume than diets that were left in a conical shape. The camera system estimated TMR volume with high accuracy across replicates (R² = 0.86). In Exp. 2, cows were introduced, and a near-real-time collection component was integrated to determine consistency of system volume estimation as TMR was consumed over time. The system detected changes in volume at the initiation and cessation of each meal bout as well as across 2 h sampling time points. Volume estimates were consistent (±2 L) over time within both individual meal bouts and each 2 h sampling time point. The results from the current study indicate that stereo vision camera systems hold some promise for estimating TMR intake in a dynamic barn environment. Further study is warranted for using stereo vision to predict TMR volume across diets and environmental conditions.