Article Preview
Top1 Introduction
With the rapid advancements of computational power and technical ability, the use of computers in public security has become a key research area in computer vision. A variety of applications including robotics, intelligent transportation, and video surveillance have become the main directions of visual intelligent analysis that have been extensively investigated (Alfaro, Mery, & Soto, 2016; Chaturvedi & Srivastava, 2016; Kulkarni, Evangelidis, Cech, & Horaud, 2015; Wang, Feng, & Yan, 2015; Wu & Ji, 2015; L. Zhao, He, Cao, & Zhao, 2016). Pedestrian detection has recently attracted much attention and significant progress has been made. However, pedestrian detection in natural scene images for behavior analysis remains an open problem.
The implementation of pedestrian detection is typically performed in one of two ways: the detection approach based on predefined features (shallow learning method), and detection methodology based on self-learning features (deep learning method). The pedestrian detection approach based on shallow learning method applies the predefined feature representation algorithm to extract pedestrian features, while the classifier classifies the objects according to the similarity between the features of objects and the pedestrian model obtained in the training process. This approach is regarded as the classical machine learning method, and is still applied in many commercial application scenarios. At the same time, the pedestrian detection method based on deep learning does not involve a predefined feature representation model. The method has the ability of the self-learning, by which it may learn the pedestrian feature representation model in the training process, and it has the ability to classify the detected objects in the testing process. The entire learning process of pedestrian detection based on the deep learning method does not require manual participation, and the detection performance is significantly improved. For these reasons, the deep learning method has received great attention in recent years.
However, with the continuous research on deep learning, its limitations have become more obvious. In order to reduce the error rate of the algorithm, the deep learning model has high requirements on the quantity and quality of hardware and training samples. This prevents deep learning from being deployed to embedded devices. In the case of insufficient training samples, the deep learning model falls into the overfitting phenomenon too early, and the error rate distance theory results will have a large gap. In the other hand, in the process of training, the requirements of hardware and training samples for shallow learning are far lower than those for deep learning. Therefore, the rapid growth of deep learning does not mean the end of shallow learning. Shallow learning method can be used as an auxiliary means to improve the detection performance of deep learning method. It is still necessary to carry out research on shallow learning method.
Within the pedestrian detection problem, feature representation has been considered one of the most important aspects. Determining image features that can efficiently and effectively discriminate pedestrians from cluttered backgrounds has been the focus of the community. Histogram of Oriented Gradients (HOG) (Dalal & Triggs, 2005a) is one of the most common hand-crafted image features for pedestrian detection. It describes the distribution of pixel gradients in a region of interest by counting the magnitudes and orientations of pixel gradients. Therefore, HOG is widely considered less complex and more comprehensible while still being effective for characterizing the target’s contour. As a consequence, many excellent detectors (Cheng, Zhang, Lin, & Torr, 2014; P. Felzenszwalb, McAllester, & Ramanan, 2008; Walk, Majer, Schindler, & Schiele, 2010; Y. Zhao, Zhang, Cheng, & Wei, 2015) combine HOG with other features to form multi-channel filters. Thus, HOG still remains a valuable research tool. One critical flaw in the method of HOG, however, is the lack of neighborhood relationships among blocks, which is important information. The algorithms mentioned above involve no discussion from neighboring relationships point of view.