Kasahara1, Takayuki Kajiwara2 and Hidetaka Nambo2, 1Industrial Research Institute of Ishikawa, Japan, 2Kanazawa University, Japan
With the growing demand for automation and efficiency in factory production, using AI anomaly detection has become increasingly important. Semi-supervised learning methods, which leverage large amounts of normal operational data, are particularly effective for this purpose. However, conventional video anomaly detection methods often struggle when applied to equipment with multiple repetitive motion patterns, generating high anomaly scores even for normal operations and leading to false positives. Moreover, while high-speed video anomaly detection systems can achieve faster processing by distributing the workload across multiple devices, they are susceptible to blind spots caused by frame loss during network transmission. To address these challenges, this study proposes a video anomaly detection method capable of accurately identifying anomalies in equipment with multiple repetitive motions and resilient to frame loss. Specifically, by taking differences in latent variable dimensions, instead of taking differences in image dimensions, it will be possible to keep anomaly scores low even for multiple patterns of operation and to indicate a high anomaly score when an anomaly occurs. This proposal can be applied not only to equipment that performs repetitive operations with a single pattern, but also to industrial equipment that performs repetitive operations with multiple patterns, such as a sorting robot that inspects goods and then places them in a basket by grade, such as grade 1 or grade 2, or a conveyor that changes delivery destinations according to destination addresses. The effectiveness of the proposed method is demonstrated through comparative evaluations with conventional approaches.
Video anomaly detection, production equipment, repetitive operations, frame lost, time information
