Anomaly Detection and Recognition in Complex Commercial Kitchen Environments

Abstract

Commercial kitchen surveillance provides important visual evidence for food-safety supervision, but automatic anomaly detection in such scenes is still challenging. Abnormal events are usually sparse and are embedded in cluttered operating environments with occlusion, low illumination, and large appearance variation. This study focuses on three representative anomalies: rat intrusion, staff smoking, and staff upper-body clothing violation. These categories cover two different recognition difficulties. Rat intrusion requires reliable tiny-object detection, whereas the two staff-related categories require the model to capture subtle local cues while also using human posture and surrounding scene context. To address these issues, this paper develops a complete framework that combines data construction and detection model enhancement. For data construction, a semi-automatic annotation-assistance workflow is built using SAM3 and Qwen3-VL-32B. SAM3 recalls candidate regions. Qwen3-VL-32B verifies candidate categories using local patches, global context, and task-specific prompts. Before manual verification, the SAM3–Qwen3-VL workflow achieves 0.928 candidate recall and 0.903 label precision. Human verification further improves the final candidate recall and label precision to 0.971 and 0.976, respectively. The complete annotation workflow requires only 38.2% of the time used by fully manual annotation. For detection, a task-oriented YOLOv11 adaptation is developed by integrating three complementary modules from prior studies. FeaturePyramidSharedConv is used to enhance high-level multi-scale context, MultiScaleGatedAttn is used to strengthen adaptive cross-layer feature selection, and DynamicScalSeq is used to reinforce the $P_{3}$ small-object branch through stacking along the scale dimension and max-response selection. Experiments on a custom kitchen anomaly dataset show that, under the default YOLOv11n setting with an input size of $640\times 640$ , the improved model achieves 0.862 mAP@0.5, which is 3.3 percentage points higher than the YOLOv11n baseline. Additional experiments further show that each adapted component contributes to the final performance, and that the framework remains effective under different MSGA placements, DynamicScalSeq variants, input resolutions, and model scales. These results indicate that the proposed method provides a practical solution for complex kitchen anomaly detection and intelligent food-safety monitoring.