Vehicle detection from complex scenes based on combination features and ELM

A new interactive detection method is proposed in order to deal with the vehicle detection problem in complex scenarios. Each target vehicle is marked by hand as the initialization to our method; the PCA-HOG and discrete moment invariants features are extracted from the manually labeled window which is considered as the bounding box of the object. Then an extreme learning machine (ELM), which is a single-hidden layer feed-forward neural network, is used to learn two kinds of features independently at the same time. At phase of detection, each detection window is checked by two trained classifiers. The experiments show that, comparing to the support vector machine (SVM), the overall performance of the proposed method is improved significantly, as the result of the great improvement by the ELM on parameter optimization and time consuming in training process without losing detection effect. As combination of features was used, the misclassified cases are reduced. This interactive detection method is suitable for vehicle detection problem under some complex scenarios.