Detection of Unusual Targets in Traffic Images Based on One-Class Extreme Machine Learning

Abstract

In traffic image target detection, unusual targets like a running dog has not been paid sufficient attention. The mature detection methods for general targets cannot be directly applied to detect unusual targets, owing to their high complexity, poor feature expression ability, and requirement for numerous manual labels. To effectively detect unusual targets in traffic images, this paper proposes a multi-level semi-supervised one-class extreme learning machine (ML-S2OCELM). Specifically, the extreme learning machine (ELM) was chosen as the basis to develop a classifier, whose variables could be calculated directly at the cost of limited computing resources. The hypergraph Laplacian array was employed to improve the depiction of data smoothness, making semi-supervised classification more accurate. Furthermore, a stack auto-encoder (AE) was introduced to implement a multi-level neural network (NN), which can extract discriminative eigenvectors with suitable dimensions. Experiments show that the proposed method can efficiently screen out traffic images with unusual targets with only a few positive labels. The research results provide a time-efficient, and resource-saving instrument for feature expression and target detection.