Research On Multilevel-based Object Detection

Object detection is one of the fundamental yet important research directions in the fields of computer vision, which can be applied in many aspects in social life. Constructing robust and efficient object detection system has great application prospect. Adopting appropriate techniques in certain levels in object detection process to form a multilevel object detection system can combine the advantages of different methods, resulting in a proper balance between efficiency and detection accuracy. Research contents in this paper include saliency detection based on spread pattern and manifold ranking, basic object detection system based on histogram of oriented gradients feature, and multi-scale object detection based on channel feature, power law and locality-constrained linear coding.Salient object detection is an important branch in object detection framework, which alone can be used as a detection method regardless of target classes, and can also be applied in preprocess stage in common object detection situations to narrow down scanning region. In this paper, a new saliency detection method based on manifold ranking and spread pattern is proposed. Accurate query information can be efficiently generated using spread pattern. Experiment results demonstrate that the method yields good detection result with concise model.Basic object detection system divides the detection process into coarse detection stage and detection window verification stage. Simple classifier using small amount of feature reject the majority of scanning windows to accomplish coarse scanning, and strong classifiers and prior knowledge strategies are used only on the scanning windows that pass the first stage. Experiment results show the effectiveness of this method to detect objects with balance between efficiency and accuracy.As for multi-scale object detection, we use channel feature and power law to construct channel feature pyramid efficiently, which is then scanned with cascade classifiers to generate coarse detection results. In scanning window verification stage, we propose to use histogram of oriented gradients together with locality-constrained linear coding method to depict the image feature, and adopt large distribution machine and feature selection strategy to classify the detection windows. Experiment results on multi-scale, multi-angle car detection dataset demonstrate the effectiveness of the proposed method.