Accuracy of a machine-vision pellet detection system |
| |
| Authors: | Kevin D Parsonage Royann J Petrell |
| |
| Institution: | Chemical and Biological Engineering, University of British Columbia, 2357 Main Mall, Rm 280, Vancouver, BC, Canada V6T 1Z4 |
| |
| Abstract: | Video cameras are employed on net pen fish farms for monitoring food pellet levels near the cage bottom. Herein, the accuracy of a new machine-vision system for the identification of a feed-wastage event and its response time are reported. Research involved novel tests and definitions, video footage recorded under different stocking and environmental conditions, and numerical filters to reduce the effects of misclassifications and patchiness of the pellet wastage record on overall system accuracy. Single-frame food pellet detection accuracy was based on the difference between computer-generated and actual frame counts of pellets greater than 30 pixels in size. A pellet wastage event was defined as a median of three or more visible pellets per frame. During tests on still video frames from different feeding events, the system missed 0.1±0.1 to 1.3±0.4 pellets frame?1, and miss-classified as pellets 0±0.1 to 2.3±0.5 objects frame?1 (n=264). Most missed pellet images were on top of fish images. Waste matter accounted for 65% of the misclassifications, while particles associated with poor camera maintenance accounted for approximately 24%. The average response time to a pellet wastage event was 5.1 and 11.4 s for sampling rates of 2 frames s?1 and 0.5 frames s?1, respectively (n=20 different pellet wastage events). The longest response time (26 s) occurred when fish were amassed against the camera and/or covering pellets. Using appropriate camera lens, camera positioning and/or ‘warning’ software can address fish-interference problems. |
| |
| Keywords: | Feed-wastage Video cameras Fish feeding Numerical filters Imaging analysis Automatic feeding Fish Net pens Aquaculture |
| 本文献已被 ScienceDirect 等数据库收录！ |
|
