Research On Methods Of Object Detection Based On Cascade Classifier

Object detection is one of the key problems in computer vision research which is widely used in areas such as human computer interaction, intelligent driving, and video surveillance. Cascade classifiers play important roles in many object detection systems because of its accuracy and great efficiency. This thesis researches on a subtype of cascade classifier, called soft-cascade classifier, and its application in object detection problems, the main content of this thesis is as follows:1. We discuss an algorithm, called soft-cascade transform which transform existing additive classifier into a much faster soft-cascade classifier without performance loss; We apply this method to additive kernel SVMs, turn them into an efficient classifier called Soft-Cascade Additive Kernel SVM(SCAK-SVM) which is suitable for object detection tasks. We research on a method called inverse feature mapping which can be used together with existing explicit feature mapping methods to speed up the training phase of additive kernel SVMs without complicating the classification phase.2. We research on the feature extraction methods which is suitable for SCAK-SVM, then combines ChnFtrs and SPHOG features to get an extended version of SPHOG, we call it Multi-level Grid Feature, this feature has good detection performance; We use a thread pool to parallelize the window searching process and increase the detection speed.3. We design a complete object detection system and discuss the key steps in system design, including parameter selection for SVM, method for cropping training images,bootstrap training strategy,image pre-processing and non-maximum surpression algorithm.We implement an experiment platform for object detection based on Matlab GUI, its functionalities includes data labeling,detector training and performance evaluation.