A vision-based system for monitoring scan-avoidance and anomalies at retail checkouts

Point-of-Sale (POS) scan-avoidance is a critical source of retail shrinkage, demanding automated trajectory monitoring solutions. However, the development of reliable automated systems is hindered by the scarcity of large-scale real-world datasets, extreme variations in product scale, and the highly dynamic motion characteristics of checkout handling. To address these challenges, this paper proposes GoodTrack, an end-to-end framework for detecting scan-avoidance anomalies. First, we establish RetailGoods , a large-scale real-world dataset constructed via a semi-automated pipeline utilizing the Segment Anything Model 2 (SAM2). Second, we propose an improved algorithmic framework leveraging enhanced You Only Look Once version 8 (YOLOv8) and ByteTrack architectures. The detector integrates the proposed Spatial Separable Star (S-Star) and Mixed-Aggregation Star (Man-Star) modules together with a Lightweight Shared Convolutional Detection (LSCD) Head to address varying product scales. The tracking module employs an Interacting Multiple Model Kalman Filter (IMM-KF) to manage erratic motion and incorporates Generalized Intersection over Union (GIoU) for robust data association in dense scenes. Compared to the YOLOv8 nano baseline, our improved detector reduces model size by 23.33% and computational load by 15.85%, while achieving 90.93% precision and 91.11% mean average precision at an Intersection over Union threshold of 0.50 (mAP@50). The tracker achieves 78.26% Multiple Object Tracking Accuracy (MOTA) and 76.81% identity F1 score (IDF1) on our real-world dataset. Finally, a large-scale validation involving 632 POS-integrated orders identified 26 confirmed scan-avoidance incidents, demonstrating the practical effectiveness of GoodTrack for intelligent loss-prevention monitoring and decision support.