| Multi- view object detection is important and fundamental in the computer vision field. In this domain, Multi-view object detection research involves wide application, including security surveillance, autonomous driving, and robotics, etc. There are two important problems for multi- view object detection: building multi- view object model effectively and localizing object of variable views rapidly. Recently, as a new valid and robust feature representation, sparse representation theory has attracted huge attention and been widely used throughout many field. Based on sparse representation theory, this paper approaches multi- view object detection via global representation, local representation and multi-scale representation. The major contributions of this dissertation are as follows:1. Since the training sample of sparse representation based multi- view is large and unable to reflect the difference of variable views,a meta-samples sparse representation based algorithm for multi- view object detection is proposed. A set of meta-samples are firstly extracted from the training samples of multi- view object to make up the dictionary, and then each sub- image was taken as a linear combination of a set of meta-samples. Then, taking advantage of the difference between representations of target and background sub- images, sparse concentration index can be used to describe target distribution in the image. As a result, a simple threshold can easily reveal target class. The experimental results show that the algorithm improves detection performance as well as speed.2. In view of the problem of partial occlusion and information missing in c urrent multi- view object represent model, a s upervised shared dictionary learning algorithm for multi- view object detection is proposed. Firstly, extract the shared features from training samples of multi- view object via NMF and learn a shared dictionary. Upon that, detect the object and estimate the view information with a discriminating function defined on the representation coefficients. Negative samples are considered as the same as the positive multi- view samples, and labels is added to the dictionary. The experimental results show that the algorithm is robust to occlusion and performs well on the detection of the mid-view object.3. In order to combine the global information and local information of multi- view object, a novel multi-scale sparse representation algorithm is proposed to represent the multi-scale information of multi- view object. Firstly, it extracted local features of multiple scales, and learned dictionaries of each scale by using sparse coding algorithm. Then sparse code the feature via learned dictionaries, spatial pools the features and contrast normalization. At last, concatenate the pooled feature of different scales to form the holistic. By utilizing the complementary relationship of features on different scales, the experimental results show that it obtains better performance on the multi- view object detection under different scale and complicated background. |
