Meet the Team that Won the SP Cup in IEEE ICASSP 2020

Q2. Congratulations on winning the contest! We’d be happy to hear about your solution on anomaly detection.

Most of the other teams have proposed solutions based on deep neural networks. Such solutions have recently become very popular, but they have several drawbacks, especially when dealing with a small amount of data, such as in this competition. Our solution stands out as it takes a different approach. We embed the noisy high-dimensional multimodal data in a low dimension representation and utilize the correlation between sensors as an indicator of abnormality. First, appearance features are extracted from each image by a pre-trained ResNet-18 deep neural network. Then, appearance and motion features are fused into a multivariate time series. Each time window is embedded in a known cone manifold, using a Gaussian kernel function. The resulting kernel matrix captures the pairwise similarity between each pair of windows. We then exploit the geometric properties of the reduced dimension manifold and use its intrinsic distance as a measure of abnormality. The proposed method is attractive as it is fully unsupervised, data agnostic, and noise-robust. As an example, we have shown that our solution can not only detect abnormal events of different types in the competition dataset, but also in another inherently different dataset.

Block diagram of the anomaly detection technique by Yair, Pavel and their team.