Key points are not available for this paper at this time.
The recognition of worker actions in pig carcass processing and the statistical analysis of their working hours are key technologies for the intelligent transformation of slaughterhouse production lines. These factors directly affect operational costs, economic efficiency, and the stability and sustainability of production. To address critical challenges such as the poor generalisability of action recognition models, low recognition accuracy, and inaccurate work hour statistics, a pig carcass processing action recognition and work hour statistics method based on Pig-Carcass Temporal Shift Module (P-TSM) is proposed. Firstly, seven key actions in the pig carcass processing workflow performed by slaughterhouse employees are analysed. Based on this analysis, a technical framework for accurate action recognition and precise working hour calculation using RGB video is developed. Two datasets, Porcine Carcass Splitting Action Dataset (PCS-AD) and Porcine Carcass Splitting-UCF101 Fusion Dataset (PCS-UCF108), are then constructed to validate the model's generalisation capability. A P-TSM-based action recognition model is established, incorporating a local temporal shift module, a label smoothing loss function, and an adaptive multi-branch network. These components enhance the model’s capacity to process temporal information and improve its generalisation performance. Based on the recognition results, and taking into account the operational sequence and working status of employees, an Action Termination Threshold and a Workpiece Process Start/Stop Determination Algorithm are designed. A method for calculating work hours based on pig carcass processing actions is subsequently proposed. Experimental results show that the P-TSM model achieved a Top-1 accuracy of 99.11% on the PCS-AD dataset and 93.85% on the PCS-UCF108 dataset. Compared with the Temporal Shift Module (TSM), the P-TSM model improved Top-1 accuracy by 3.61 percentage points and Top-5 accuracy by 0.46. It also outperformed other mainstream action recognition algorithms, enabling high-precision recognition of worker actions in pig carcass processing. On-site testing demonstrated that the overall accuracy of the proposed method for working hour statistics reached 96.67%, providing solid technical support for the intelligent upgrading and production management of slaughterhouse lines.
Xianbin et al. (Sun,) studied this question.