SaigeVAD® Video Anomaly Detection System

Even the best machines make mistakes—assembly parts may be misaligned or even fall off, and some crucial pieces may be missing entirely. Failure to immediately detect such manufacturing assembly errors can lead to serious quality control issues, and cause long manufacturing downtimes later while the source of the problem is traced. The SaigeVAD® video anomaly detection system is designed to automatically detect such manufacturing assembly errors. Like SaigeVision®, SaigeVAD® is based on our proprietary deep learning technology to quickly and reliably detect all anomalies. Using only unlabelled training data—a continuously recorded video sequence of normal manufacturing assembly operations, for example— SaigeVAD® can be quickly and efficiently trained to not only detect all anomalies, but also to classify which of these are acceptable or unacceptable.

1. Robust to changes in lighting, video angles, the assembly environment, and the assembly process.

  • Unlike rule-based video anomaly detection systems, SaigeVAD® can detect anomalies under a wide range of lighting conditions, video camera angles, and changes in the assembly environment and the assembly process. A network trained only once can still reliably detect anomalies under small changes in the environment.
  • For larger changes in the environment, SaigeVAD® automatically calculates the output value shift due to these changes, and adjusts its pre-trained model to the new environment; no retraining is required.

2. No labelling required.

  • Labelling data is time-consuming and costly, especially video data. SaigeVAD® only requires video of normal manufacturing assembly operations. This can include a wide range of anomalies that are acceptable. For example, some parts can be attached to each other at a range of acceptable angles and alignments; these should be regarded as acceptable anomalies, and not flagged for rejection.
  • Even with only a few instances of such acceptable anomalies, SaigeVAD® can learn and classify anomalies accurately and reliably.

3. Fast, lightweight, and robust video anomaly detection optimized for manufacturing assembly

  • Manufacturing assembly requires both accuracy and speed: Anomalies must be detected in tenths of seconds, and over 99% accuracy must be achieved while ensuring that false negatives stay below 1%.
  • With their large networks, and similarly large data and labelling requirements, video anomaly detection systems developed for surveillance and security will not work for manufacturing assembly operations.
  • Employing our proprietary deep learning technology optimized for the repetitive nature of manufacturing assembly operations, SaigeVAD® deliver real-time, robust performance using light networks that can be easily and rapidly trained.