Overview of Pedestrian detection

From: http://blog.csdn.net/carson2005/article/details/8316835

I declare that this article is translated, summarized, and summarized after I have read several summary articles about Pedestrian detection. This helps you understand the development trend of Pedestrian detection. At the same time, it also provides some ideas for friends who intend to study Pedestrian detection.

Pedestrian detection history:

In the early days, segmentation, edge extraction, motion detection, and other methods were mainly used in static image processing. For example:

(1) Global template method represented by Gavrila: Based on the contour layered matching algorithm, nearly 2500 profile templates are constructed to match pedestrians to identify travelers. In order to solve the speed reduction problem caused by the large number of templates, a layered search policy from coarse to fine is adopted to accelerate the search speed. In addition, when matching, the similarity between the two is measured by calculating the distance transformation between the template and the window to be detected.

(2) Local template method represented by broggi: uses two-value image templates of different sizes to model human heads and shoulders, the edge image of the input image is compared with the binary template to identify pedestrians. This method is used in the ARGO Smart car developed by Parma University in Italy.

(3) optical flow detection method represented by Lipton: Calculate the residual optical flow in the moving area;

(4) motion detection method represented by heisele: extract the characteristics of pedestrian leg movement;

(5) Wohler-represented neural network method: Build an adaptive time delay neural network to determine whether it is a human motion image sequence;

The above methods are characterized by slow speed, low detection rate, and high false positive rate.

Current Situation of Pedestrian detection:

There are two types:

(1) Background-based modeling method: This method divides the foreground, extracts the moving target, and then further extracts features and classifies them; this method is not robust and has poor anti-interference ability in scenarios such as rain, snow, windy weather, shaking leaves, and dark lights. The model of the Background Modeling Method is too complex and sensitive to parameters.

(2) Statistical Learning: A pedestrian detection classifier is constructed based on a large number of training samples. The extracted features generally contain information such as the gray scale, edge, texture, shape, and gradient histogram of the target. The classifiers include neural networks, SVM, and AdaBoost. This method has the following difficulties:

(A) different attitudes and costumes of pedestrians;

(B) The distribution of extracted features in the feature space is not compact;

(C) The performance of the classifier is greatly affected by the training samples;

(D) The negative samples during offline training cannot cover all real application scenarios;

Although the pedestrian detection method based on statistical learning has many shortcomings, many people still focus on this. A typical representative is the hog + SVM Pedestrian detection algorithm (histograms of Oriented gradients for human detection, navneet dalel, Bill triggs, cvpr2005) published by Dalal, a French researcher in cvpr 2005 ). Hog operator is introduced in detail in another blog (http://blog.csdn.net/carson2005/article/details/7782726), while hog + SVM Pedestrian detection algorithm has been implemented by opencv, the author of a blog has a detailed process introduction and reference code: http://blog.csdn.net/carson2005/article/details/7841443

Considering the advantages and disadvantages of Background Modeling and statistical learning, researchers have proposed to combine these two methods for fast and accurate Pedestrian detection. A typical system is as follows:

Note: from this article: robust Pedestrian detection algorithm based on scene model and statistical learning; Journal of automation; July April 2010;

In this paper, the author uses GMM for background modeling, uses Haar-like features to describe pedestrian features, and uses AdaBoost cascade structure as a classifier. An Improved weak classifier selection algorithm is proposed to enable the selection of weak classifier and the re-training of the classifier to be completed in about 10 minutes.

Current typical pedestrian detection directions

(Note: State of the art. I do not know how to translate this word. I am writing "typical directions" for the time being. Please tell me if you know it );

Pepageorgios and others were the first to adopt Sliding Window for pedestrian detection. They used SVM and multi-scale Haar wavelet over-complete basis for pedestrian detection. Based on this idea, Viola and Jones use integral graphs to achieve fast feature computation, and use a cascade structure for efficient User Detection. At the same time, automatic feature filtering using Adaboost algorithm. These ideas constitute the cornerstone of today's Pedestrian detection operators.

Inspired by the sift operator, Dalal and triggs proposed the use of histogram of gradient (hog) features in pedestrian feature description, experiments show that hog is more informative than grayscale-based features. Shahua and others have also proposed a similar method to portray pedestrians. Since then, the hog-Based Variant methods have increased dramatically, and all these variants have adopted some ideas of the hog operator to a certain extent. Shape features are also an effective feature description method for pedestrian detection. Gavrila and philomin use the Hausdorff Distance Transformation and a layered template matching method to quickly detect pedestrians. Wu and nevatia use a large number of line segments and curves to form a feature called "edgelet" to express shape features locally. Researchers also used the boosting method to learn the detection operators of the head, trunk, leg, and whole body. Similarly, some researchers have proposed a feature called shapelets, which depicts shape features based on a gradient of local image regions (patches.

Motion is another important clue in pedestrian detection. However, in the case of camera motion, effective use of motion features is a challenging topic. When the camera is fixed, Viola and others propose that the Haar-like features of different images can be computed to improve the performance. When the camera is not fixed, the motion classification needs to be decomposed. Dalal and other people use the light flow field to carry out statistical modeling of the movement inside the image, and then make certain Motion Compensation in the partial area of the image.

For a single feature, there is no feature descriptive operator that is more effective than the hog operator. Of course, other features can be combined with hog features to complement them. Wojek and Schiele research found that by combining Haar-like, shapelets, shape context, hog features, it will be more effective than any other independent feature description operators. On this basis, walk et al. considered color self-correlation and the previously mentioned motion features. Similarly, Wu and nevatia combine hog, edgelet, and covariance features. Wang et al. proposed to combine the texture features and the hog Operator Based on HSV. In addition, the SVM classifier was improved to make it more suitable for occlusion. Of course, some people also propose to combine the partial tri-value pattern (a variant of HSV), color information, and implicit segmentation with hog. Of course, the above method is better than the pure hog method, performance has been improved to a certain extent.

Dollar et al. expanded on the basis of Viola and Jones and proposed Haar-like feature extraction on multiple channels, including Luv color channel, gray level, gradient amplitude, etc, this method is a hodgedge of multiple features. Tuzel and others use the covariance matrix of special local features as a feature description method. In addition, researchers have focused on "How to effectively utilize the vast feature space ". Therefore, Feature Mining was proposed by researchers to train boost classifiers using various strategies, including the maximum descent method.

Pedestrian detection: an evaluation of the state of the art, PAMI, 2012; a friend of mine can download: http://download.csdn.net/detail/carson2005/4904088