Due to unknown flesh and bone boundaries in the hind legs of sheep carcasses, variable size and visibility constraints, the robot autonomous segmentation accuracy is low and easy to be blocked. An adaptive segmentation control method was proposed for the hind legs of sheep carcasses, and the segmentation test of sheep carcass hind legs was carried out to verify it. The method was centred on contact state perception and effectively extracted contact type features, contact abnormality features and contact direction features. LSTM-FCN deep spatio-temporal neural network was constructed to identify contact types, constructing deep self-coding network to estimate contact anomalies, and using principal component analysis to detect the main contact directions to achieve multimodal sensing of contact states. The robot imitated and learned human manipulation skills through dynamic motion primitives, and incorporated contact state sensing information to achieve adaptive adjustment of joint motion. The experimental results showed that the recognition accuracy of LSTM-FCN model on the validation set of sheep carcass hind leg segmentation was 98.44%, with a high recognition accuracy. The DAE model can better estimate the contact anomalies of the validation set samples and distinguish different contact states. Robot conducted practical segmentation tests based on adaptive segmentation control method. Compared with the control group, the maximum segmentation force was decreased by 29N and the maximum torque was decreased by 7N·m, proving the effectiveness of the method. The average maximum residual meat thickness was 3.6mm, the average segmentation residual rate was 4.9%, and the segmentation residual rate showed a negative correlation with the quality of sheep carcasses. It proved that the method had good generalization and accuracy. And the overall segmentation effect was good, meeting the requirements of sheep carcass hind leg segmentation.