Pedestrian Detection Based On Deformable Part Models And Histograms Of Sparse Codes Features

Human detection has been a research focus in the field of computer vision and pattern recognition, as times progressed, people’s life is becoming more and more rich and colorful with the rapid development of modern high-tech, many fields ar involved in human detection, such as intelligent entrance guard system, the subway station. Both the human body detection system and the human body detection technology is portable to other platforms, human detection improved will naturally improve human detection system. So far, there are many human detection methods, but because the human body itself has more flexible variability and human environment is complex, human detection is still a challenging research topic. In recent years, the deformable component model shows more superiority in human detection, it can improve the efficiency of human detection using the relationship between the body’s overall information and part information, even the complex human body target or in the complex environment, the model can exhibit very good performance.This article focuses on the study of target detection model and a description of the image features operator. We define these models using a grammar formalism. In this formalism compositional rules are used to encode models that can range in complexity from simple rigid templates to rich deformable part models with variable structure. This article is to explore this idea along, gradually enrich pedestrian target detection model. By building and training complex models structures to improve pedestrian detection performance, And PASCAL VOC Challenge data sets are used to validate the models.At the same time build a richer model, we use existing training data, the training data are usually specified labels, such as object bounding boxes, compared to the grammar model, these labels can’t indicate the deformable components, is a “weak” label. This article introduces a kind of discrimination learning ability of weak label to label data sets for training. Compared with traditional training methods have greatly improved performance.The HOG features, heavily engineered for both accuracy and speed, are not without issues or limits. They are gradient-based and lack the ability to directly represent richer(and larger) patterns. We provide an affirmative answer by proposing and investigating a sparse representation for object detection, Histograms of Sparse Codes(HSC). We compute sparse codes with dictionaries learned from data using K-SVD, and aggregate per-pixel sparse codes to form local histograms.In order to apply the study to the actual application, the pedestrian detection target system must not only be accurate, but also and fast. To solve this problem, we use the cascade detection framework. By using the method of selecting cascade thresholds, the detection rate is increased more than 4 times.