Validation of a real-time location system for zone assignment and neighbor detection in dairy cow groups

Highlights

• Professional calibration of RTLS based on tachymetric surveying recommended.

• Low tracking data quality sufficient for individual-level analysis.

• High tracking data quality enables social behavior research.

• Recommendations provided for evaluating and testing an indoor tracking system.

• Use of additional automatic devices is beneficial for evaluation.

Abstract

Tracking data are increasingly used for studying social behavior in dairy cows but guidelines are not available regarding the appropriate system setup and data quality. In this study, we investigated the effects of system calibration, as well as data filtering and smoothing methods on the detection of the location and neighbor preferences of dairy cows in a barn using an ultra-wideband real-time location system (RTLS). We followed a group of 15 or 14 cows during two periods: period 1 with RTLS calibration using a laser distance meter; and period 2 with RTLS calibration using professional surveying technology. During both periods, tracking data were collected every 1.73 s for 3 days for all cows in the group together with data from an electronic feeder system. We performed continuous video analysis of cow locations for 1 day in both periods. We compared the effects of four different filtering and smoothing methods (i.e., Kalman filter, sliding window, jump filter, and median filter) on data preparation using raw tracking data. We assigned specific areas (i.e., lying stalls, walking alley, brush area, feed, and water bins) to the measured X-Y coordinates and compared them with the video-based zone assignments. Sensitivity and precision were calculated to evaluate the quality of the zone assignments corresponding to the different RTLS calibration setups and the effects of filtering and smoothing methods. When the RTLS was accurately calibrated these methods did not result in further improvements. The zone assignments agreed well for the video and all of the prepared tracking data at the lying stalls in both periods, and the agreement was good during period 2 at the feed bunk (sensitivity and precision >0.85). When using zone-based and distance-based approaches for detecting neighbors at the lying stalls and feed bunk based on the prepared tracking data we found high correlations (>0.9) with the neighbor information based on video and electronic bin data in period 2. However, the zone-based approach resulted in the lowest mean absolute error. The quality of the tracking data appeared to be especially important when detecting cows in small areas and with short visit durations (e.g., feed bunk). Our results highlight the importance of accurate RTLS calibration setup and the quality of tracking data when inferring dairy cow behavior based on it.