Real-Time Working-Point Detection and Estimation for an Evisceration Robot Using a Single Time-of-Flight Camera and a Shared-Backbone Multi-Task Network

Evisceration is a slaughter stage with strong demand for automation. The robot must detect four working-point classes and estimate each working point in pixels. We compose the network input from the luminance and depth images of a single Time-of-Flight (ToF) camera, avoiding the registration error of RGB-D cameras, and propose a retroreflective spherical-marker target for hand–eye calibration; the calibration procedure built on it was 3.1 and 3.3 times more repeatable in translation and rotation than a grid-target procedure. We extend the RF-DETR-L detector into a 43.9 M-parameter multi-task network in which a keypoint branch initialized from the detection outputs shares the DINOv2-S backbone. On five carcasses from the same facility not used for training (500 images, three runs), the network reached a detection AP of 68.6 and a keypoint AP of 73.3, comparable to the serial detector–pose-estimator pipeline it replaces (68.3 and 73.2), with a mean working-point error of 1.83 vs. 2.01 px. With TensorRT FP16, its network-stage latency was 2.34 ms (plus 3.80 ms resize), compared with 4.11 ms for the serial pipeline (3.40 ms batched), and stayed nearly constant when up to sixteen objects were forced (serial: 17.22 ms); a single-stage YOLOv8l-pose gave similar AP and network-stage latency.