Study Of Image-based Human Detection Algorithm

In recent years, human detection is a hot research topic of computer vision; it has broad applications in the field of image retrieval, intelligent transportation, intelligent video surveillance, advanced human-computer interaction etc. Human detection is a challenging problem due to the uncertainty of the appearance, shape, the complexity of the scenarios, mutual occlusion causes etc. Although the human detection technology has been researched for many years, but there is not a common, robust, accurate and real-time human detection algorithm. At present, human detection algorithm exists two problems:(1) the detection rate is not high, making it difficult to meet the accuracy requirements of applications;(2) the detection speed is slow, making it difficult to meet the requirements of real-time applications.In order to solve problems in human detection, this paper proposes a series of methods to improve the detection rate and detection speed. This paper proposes a human detection method based on cascade FIK-SVM classifier, which effectively improves the detection rate of human detection. In order to improve the speed of muti-scales detection, we use muti-order integrable features, image pyramid acceleration and multi-pose learning boost to optimize the speed of detection and multi-scale feature extraction. This paper also examines the proplem of human detection with scene context, and proposes a novel method which uses scene information to improve the detection rate. The main contributions of this paper are follows:(1) This paper proposes a human detector based on cascade FIK-SVM classifier, which improves the detection speed of human detection algorithm. We propose a simplified LBP texture feature, and combine this feature with variable-size blocks of HOG feature to get feature vector. The feature not only has high discrimination, but also has high extraction speed by using integral histogram. In order to improve the discrimination of traditional linear SVM constructed weak classifiers, we propose a method which using FIK-SVM hyperplane to build weak classifiers. Weak classifiers are cascaded to form a strong classifier by using Adaboost. Detection speed of our method is close to other methods with cascaded weak classifiers, while the detection accuracy is close to other methods with strong features and strong classifier. The classifier effectively improves the detection rate at the same time ensure rapid detection speed.(2) This paper proposes a real-time human detection algorithm based on multi-order integrable features pool and multi-pose learning boost, which effectively improve the speed of human detection algorithm, while also improve the detection rate. We propose multi-order integrable features and use these feature to build weak classifiers. Because all these features can be computed by using integral image, so the feature extraction process is very fast. Weak classifiers are cascaded to form a strong classifier by using multi-pose learning boost. We simultaneously train several strong classifiers, while the response is given as the maximum response of all the strong classifiers. Thus, a detection window is classified as positive if a single strong classifier yields a positive score and negative only if all strong classifiers yield a negative score. This paper proposes a method for multi-scale detection optimization. This method can avoid constructing a finely sampled image pyramid without sacrificing performance. This is based on the principle that for many features, including gradient histograms, the feature responses computed at a single scale can be used to approximate feature responses at nearby scales.(3) This paper proposes a super-pixel clustering and contextual features pool based human detection method, which effectively uses context information to further improves the detection rate. We first split pixels into super-pixels, and then train classifier to allocate super-pixels into different scene area, and then extract contextual features in the scene area map. In order to take advantage of scene information, we randomly select contextual features and local features to build weak classifiers, then we use multi-pose learning boost cascade these weak classifiers into strong classifier. Experiment results show that this method has higher detection rate than traditional human detection algorithm and show faster speed than previous context-based human detection algorithms.