Anda di halaman 1dari 15

CROWD COUNTING USING

CONVOLUTIONAL NEURAL
NETWORK
OVERVIEW
OBJECTIVE

INTRODUCTION

RELATED WORKS

PROPOSED MODEL

SUMMARY

REFERENCES
Objective
•To counting the number of people in the images with a
help of convolutional neural network.
•To train CNN to estimate density map for the patches
of the images.
•To train Switching network to estimate suitable CNN
regressor.
INTRODUCTION
Crowd counting is the important phenomenon used for
security reason, getting the number of people in the crowd for
controlling the crowd in sudden emergencies.
This project aims to take a broader view to address crowd
counting from the perspective of semantic analysis. The
factors involving in the semantic analysis are people, heads
and their context structure.
The deep convolutional neural networks (CNN) based crowd
detection mechanism is proposed for separating the task into
subtask.
The advantage of deep learning based approaches require
huge amount of training data (more than 100 training datasets)
and to learn scale invariant representation of multiple dataset
Related works

Viola-Jones Face Detection Algorithm


The Viola-Jones face detection method uses combinations
of simple Haar-Like features to classify faces. Haar like
features are rectangular digital image features that get their
name from their similarity to Haar-wavelets .
Disadvantage:
It can detect only the few face in the images but not all
the face in image even the author apply the robustness of
face detector.
Generic Head Detector:

For an input image, a set of channels are computed. After


we detected the head and few other aspects in the images
and we can mark it and count.
Disadvantage :
Less accuracy in dense crowd
Shape-based human detection
A hierarchical part-template matching approach and
combine it with discriminative learning for building a generic
human detector.

Disadvantage
Mismatching may happen at different view point so
the accuracy will be reduced .
Viewpoint invariant approach

It is a learning based method that uses feature histogram and


feature normalization to deal with perspective projection and
camera orientation.

Disadvantage:
Hard to train
Less accuracy
Convolutional Neural Network
Convolutional networks were inspired by biological processes
(animal visual cortex). A CNN consists of an input and an output layer, as
well as multiple hidden layers. The hidden layers of a CNN typically
consist of convolutional layers, pooling layers, fully connected layers,etc.,
Proposed Methodology
Crowd Switching
image network
patches

CNN CNN CNN


r1 r2 r3

Density
map

Crowd count
Training method
Pretraining
Differential training
Coupled training

Dataset used
Shanghai-Tech
References
Huang, Siyu & Li, Xi & Zhang, Zhongfei & Wu, Fei & Gao, Shenghua & Ji, Rongrong
& Han, Junwei. (2017). Body Structure Aware Deep Crowd Counting. IEEE
Transactions on Image Processing. PP. 1-1. 10.1109/TIP.2017.2740160.
H. Idrees, I. Saleemi, C. Seibert, and M. Shah, “Multi-source multi-scale counting
in extremely dense crowd images,” in Proc. IEEE Conf. CVPR, 2013, pp. 2547–
2554
V.B.Subburaman, A.Descamps, and C.Carincotte,”Couning people in the crowd
using generic head detector”, in Proc.IEEE Conf. AVSS.2012,pp.470-475.
Y. Bo and C. C. Fowlkes, “Shape-based pedestrian parsing,” in Proc. IEEE Conf.
CVPR, 2011, pp. 2265–2272
Z. Lin and L. S. Davis, “Shape-based human detection and segmentation via
hierarchical part-template matching,” IEEE Trans. Pattern Anal. Mach. Intell., vol.
32, no. 4, pp. 604–618, 2010.
Theo Ephraim ,Tristan Himmelman and Kaleem Siddiqi, “Real time viola-jones
face detection “,in Proc.IEEE Conf.2009.
D. Kong, D. Gray, and H. Tao, “A viewpoint invariant approach for crowd
counting,” in Proc. IEEE ICPR, vol. 3, 2006, pp. 1187–1190.

Anda mungkin juga menyukai