Article Directory
Research progress
Pedestrian target detection methods are mainly divided into four categories:
1. Based on background difference/interframe difference
The key of the background difference method is to construct a suitable background. The classic method is the background method of the mixed Gaussian model.
The inter-frame difference method and the background difference method are similar in the principle of pedestrian detection. If the pixel difference between frames is greater than the set threshold, it is judged that there is a moving target.
Proposed solutions:
a. An adaptive Gaussian mixture background model based on pixel points and spatial distribution, and non-parametric estimation to build a spatio-temporal and Gaussian background model
. b. Multidimensional Gaussian mixture model + inter-frame difference method, suitable for a fixed background modeling method
c. Introduce acceleration factor and rationality feedback
in the background update d. Based on the new faster than background model model update mode
e. According to the characteristics of pedestrians in road traffic, Li Juan uses the classical Gaussian mixture model to update the parameters and estimate the background Improvements have been made in the three aspects of foreground and foreground segmentation, which solves the difficulty of
detecting the mixed Gaussian model . f. Inter-frame differences generally use two frames of images to do the difference. For example, Damien uses the inter-frame difference to obtain the background, and performs detection on the obtained background image according to the detection results. Selective update to detect moving targets.
g. Use the inter-frame difference of continuous multi-frame images
h. Combine the frame difference method with other methods
i. Combine the inter-frame difference and the background difference
2. Pedestrian detection method based on optical flow
The optical flow method mainly uses continuous video frames to calculate the motion vector of each pixel to establish an optical flow field, and then uses the continuous change of the motion vector to detect whether there is a moving target.
a. Differential image absolute value + optical flow method
b. Target detection algorithm based on spatio-temporal interest points, which solves the shortcomings of the differential optical flow method when the illumination changes greatly
. c. Optical flow method + SMV, generally the optical flow method is combined with the optical flow method. Combination of other methods
3. Pedestrian detection method based on template matching
The template matching method was first used to detect rigid objects with a fixed shape. Later, some scholars used the template matching method in pattern recognition to detect pedestrians in the video by constructing feature templates such as human contour, torso, head, and arms. Template matching method can be used to detect stationary pedestrians. It has the most extensive research and has tracking capabilities.
Solutions have been proposed:
a. Motion + edge extraction method + Gaussian mixture model to model human head contour b.
Hough transform
c. Density mapping relationship to improve the effect of pedestrian head detection
d. Layered contour template matching method, from coarse to western Hierarchical search strategy to improve the detection speed of the system
e. Fusion of LW-PGD, HOG features, HSV color features to quickly retrieve the human head and shoulders.
f. Multi-view cascaded head detection method based on Adaboost
h. Using neural network to initially recognize human contour + hair model and mean shift to occlude the non-human target image to cluster multiple human head and shoulder models.
4. Based on machine learning
This method designs representative pedestrian features (texture, color, histogram of directional gradients (HOG), Haar features, scale-invariant features (SIFT features), accelerated robust features (SURF features), etc.) and efficient feature classification Detector (support vector machine (SVM), random forest, deep learning, etc.) to achieve accurate pedestrian detection.
Solutions have been proposed:
a. Extract a sample or a histogram of gradient vectors (HOG) of each block in the detection window for training to obtain an SVM classifier.
b. Codebook background modeling algorithm extracts pedestrian foreground images, reducing the search range of the retrieval process.
c. Random forest learning method, fuzzy random forest learning method, using human Haar features and HOG for training and classification.
d. Based on Adaboost and face SURF features of the Mormeter detection method + YCbCr space to establish a skin color model for skin color verification.
e. People
counting method based on Harr+Like local features + Adaboost + cascaded background differential repair algorithm
f. Adaboost+SVM combined to form a cascaded classification algorithm
g. Sequential iterative training of multiple convolutional neural networks is proposed based on the verification set A strategy for better model selection based on the accuracy rate and the stability of the classifier.
h. Deep convolutional neural network, learn pedestrian characteristics from the pedestrian library, and then employ selection search algorithm, binary normalized gradient algorithm to remove a large number of redundant windows
i. Multi-layer network construction, focusing on the analysis of the number of network layers, the size of the convolution kernel, The influence of parameters such as feature dimension on pedestrian detection, optimization of network parameters.