Pedestrian Detection Deep Learning Chapter

• Shang Xu June, human body behavior recognition based on Deep learning J Wuhan University Journal 2016414492-497
  • Introduction
  • Behavior Recognition Overall process
  • Foreground extraction
  • Behavior Recognition Process
  • Experimental analysis
• Computer Engineering and application of pedestrian detection based on deep convolutional neural network in Rui 2015
  • Introduction
  • The structure and characteristics of convolutional neural network
  • Pedestrian detection convolutional neural network structure
  • Experimental comparison Summary
• Zhang Yang Pedestrian detection method based on belief degree network classification algorithm J Computer Application Research 20163302

In general, most of the browsing is done.

Shang, Xu June and so on. Human behavior recognition based on deep learning [J]. Journal of Wuhan University, 2016,41 (4): 492-497.0 Introduction

目前研究行为识别的方法一般分为基于模型方法和基于相似性度量的方法，前者首先建立某种准则，然后从运动图像序列中提取目标的外形、运动等特征，根据所获得的特征信息，通过人工或半监督的方法来定义正常行为的数学模型。而基于相似度量的方法考虑到人体行为难定义、易发现的特点，避免显示定义人体行为的数学模型。其基本原理是自动从运动图像序列数据中学习各种人体行为，根据学习结果判断测试视频中的行为类型。本文提出了一种基于深度信念网络（deep belief networks）的人体行为识别方法。

1 Behavior Recognition Overall process

The flowchart is as follows:

The left branch is the model training, and the right model is the recognition process.

2 Foreground extraction

At present, the target detection methods mainly include background subtraction, optical flow method and time difference method. In order to realize the self-adaptability and real-time performance of detection, this paper chooses the background subtraction based on Gaussian mixture model, which is simple, fast in operation and adaptable to background changes. (The time difference method is generally more difficult to extract the complete moving target, and it is easy to create voids inside the moving target.) The calculation of optical flow method is relatively complex and the noise-resisting ability is poor. Specific mathematical operations, the blog is no longer detailed, is the effect:

(The background should not be so simple ~ ~)

3 Behavior Recognition Process

The deep learning-related content used in the behavioral recognition process is no longer detailed here, and there is a separate blog discussion behind it. Can learn from this http://blog.csdn.net/zouxy09/article/details/8781396 first
The models used in deep learning include automatic encoder, sparse coding, depth belief network, and so on, this paper chooses DBNS model.
In a deep neural network, any two adjacent hidden layers constitute a restricted Boltzmann machine (Restricted Boltzmann machines RBM), and the depth belief network is a probabilistic model with multiple hidden layers, each of which acquires highly correlated correlations from the previous hidden layer, Can be seen as the accumulation of multiple RBM, each low-level RBM output as input data used to train the next RBM, through greedy learning to get a set of RBM, this group of RBM can constitute a DSNs, such as

This paper chooses greedy layered training algorithm. In the process of greedy learning, the idea of Wake-sleep algorithm is adopted. The learning process is as follows:

4 Experimental Analysis

In this paper, the relationship between the number of hidden layers and the frequency of iteration and the training error is compared.
The results obtained by comparison with other literatures showed better.

Computer Engineering and application of pedestrian detection based on deep convolutional neural network in Rui 2015

The process of general convolution neural network application can be familiar with this paper.

0 Introduction

Compare with HOG+SVM or adaboost, get better results

1 convolutional neural network structure and characteristics

convolutional Neural network explanation, the following article is better
http://www.36dsj.com/archives/24006

2 Pedestrian detection convolutional neural network structure

The classical convolutional neural network can not complete the pedestrian detection task effectively, the network depth, convolution kernel size and the final feature dimension are the main factors affecting the result. Therefore, it is necessary to redesign the structure of convolutional neural network for the specific characteristics of pedestrian detection problem. Consider the question mainly:
1. Impact of convolutional nuclei. Convolution kernel is the most characteristic part of convolutional neural network model, which can be understood as the model expression of the sensing field in biological vision. Its nature directly determines the quality of feature extraction, the speed of network convergence and so on. The size of the convolution nucleus determines the size of the field of sensation, the feeling of the wild is too large, the characteristics of extraction beyond the convolution nucleus of the expression range, and the feeling is too small, you can not extract effective local features. Therefore, the convolution kernel size has a critical impact on the performance of the entire network.
2. By increasing the number of layers in the network, its characteristic information expression ability is gradually enhanced, but too many layers will lead to the network structure is too complex, the training time increases, prone to the occurrence of overfitting phenomenon
3. Influence of classifier input feature dimension
Based on the above analysis, the pedestrian detection convolutional neural network is redesigned, and the above thought and network structure parameters are verified by experiments. Finally, the deep convolutional neural network structure is divided into 7 layers, the convolution kernel size is 9x9, and the output feature dimension of the hidden layer is about .

3 Experimental Comparison Summary

Still the same old, the experiment result is very good, the method is very good ...

Zhang Yang Pedestrian detection method based on belief degree network classification algorithm [J] Computer Application Research 2016,33 (02)

Pedestrian Detection Deep Learning Chapter