中国科学院软件研究所机构知识库

In recent years, using computer in the field of intelligence has drawn a widespread concern. Object detection, as one of the key technologies, has become a hotpot of pattern recognition and computer vision, in order to meet the requirements of human-computer interaction, intelligent surveillance system and content-based multimedia retrieval. However, because of changing illumination and perspective, occlusion and real-time detection, object detection is still a very challenge research.

First, this paper makes in-depth research and analysis of papers in object detection of recent years, summarizes the performance of feature extraction and machine learning algorithms through making a compare of computational complexity, accuracy, robustness and applicability of different algorithms, and concludes the scenarios of different algorithms.

Secondly, this paper proposes a variety of evaluation parameters and a feature evaluation system based on cluster analysis, to achieve quantitative analysis of the performance of different feature. After analyzing a large number of samples, object priori-knowledge is applied to the process of feature extraction. Moreover, under the principle of feature evaluation herein, this paper proposes two feature extraction operators, one is based on histogram of oriented gradient of stable area and the other is based on the stable region haar-like feature.

Finally, this paper proposes stable region feature based on multi-simple feature, which simulates the perception process of human visual system. Then, the extraction and optimization of stable region feature, optimization of multi-scale and multi-object detection are researched. As shown in the experiment of pedestrians under natural conditions, algorithm proposed by this paper can perform better both in accuracy and in efficiency than those similar algorithms.