violent video detection based on mosift feature and sparse coding long xuchen gongjie yangqiang...
TRANSCRIPT
![Page 1: VIOLENT VIDEO DETECTION BASED ON MoSIFT FEATURE AND SPARSE CODING Long XuChen GongJie YangQiang WuLixiu Yao Aluno: Rómulo Ramos Avalos](https://reader036.vdocuments.net/reader036/viewer/2022062700/552fc174497959413d8ef188/html5/thumbnails/1.jpg)
VIOLENT VIDEO DETECTION BASED ON MoSIFT FEATURE AND SPARSE CODING
Long Xu Chen Gong Jie Yang Qiang Wu Lixiu Yao
Aluno:Rómulo Ramos Avalos
![Page 2: VIOLENT VIDEO DETECTION BASED ON MoSIFT FEATURE AND SPARSE CODING Long XuChen GongJie YangQiang WuLixiu Yao Aluno: Rómulo Ramos Avalos](https://reader036.vdocuments.net/reader036/viewer/2022062700/552fc174497959413d8ef188/html5/thumbnails/2.jpg)
INTRODUÇÃO Os momentos de detecção de violência no vídeo está métodos da descrição do espaço-tempo local das consultas nos vídeos comumente aplicado. No entanto, esses descritores não são suficientemente discriminativos.
Este trabalho usa Movimento SIFT (MoSIFT) para a descrição de nível baixo do vídeo, Kernel Density Estimation (KDE) para seleção de recursos, e finalmente, usando Sparce Coding Scheme para obter melhores resultados na discriminação de recursos.
![Page 3: VIOLENT VIDEO DETECTION BASED ON MoSIFT FEATURE AND SPARSE CODING Long XuChen GongJie YangQiang WuLixiu Yao Aluno: Rómulo Ramos Avalos](https://reader036.vdocuments.net/reader036/viewer/2022062700/552fc174497959413d8ef188/html5/thumbnails/3.jpg)
Dataset
1000 videos de hockey fight dataset quais 500 são violentos e 500 não. Cada clipe tem 50 quadros com uma resolução de 360x288 pixels.
246 videos cwowd violence dataset quais 123 conjunto de dados que são violentos e 123 não, com resolução de 320x240 pixels.
![Page 4: VIOLENT VIDEO DETECTION BASED ON MoSIFT FEATURE AND SPARSE CODING Long XuChen GongJie YangQiang WuLixiu Yao Aluno: Rómulo Ramos Avalos](https://reader036.vdocuments.net/reader036/viewer/2022062700/552fc174497959413d8ef188/html5/thumbnails/4.jpg)
Hockey Fight Dataset
Violence Non-Violence
![Page 5: VIOLENT VIDEO DETECTION BASED ON MoSIFT FEATURE AND SPARSE CODING Long XuChen GongJie YangQiang WuLixiu Yao Aluno: Rómulo Ramos Avalos](https://reader036.vdocuments.net/reader036/viewer/2022062700/552fc174497959413d8ef188/html5/thumbnails/5.jpg)
Crowd Violence Dataset
Violence Non-Violence
![Page 6: VIOLENT VIDEO DETECTION BASED ON MoSIFT FEATURE AND SPARSE CODING Long XuChen GongJie YangQiang WuLixiu Yao Aluno: Rómulo Ramos Avalos](https://reader036.vdocuments.net/reader036/viewer/2022062700/552fc174497959413d8ef188/html5/thumbnails/6.jpg)
Framework of the proposed violence detection approach
![Page 7: VIOLENT VIDEO DETECTION BASED ON MoSIFT FEATURE AND SPARSE CODING Long XuChen GongJie YangQiang WuLixiu Yao Aluno: Rómulo Ramos Avalos](https://reader036.vdocuments.net/reader036/viewer/2022062700/552fc174497959413d8ef188/html5/thumbnails/7.jpg)
MoSIFT Algorith
Aplica SIFT Estándar para encontrar pontos de interesse visualmente distintos no domínio espacial.
Aplica-se também um analogous histogram of optical flow. Para rejeitar candidatos com recursos insuficientes.
![Page 8: VIOLENT VIDEO DETECTION BASED ON MoSIFT FEATURE AND SPARSE CODING Long XuChen GongJie YangQiang WuLixiu Yao Aluno: Rómulo Ramos Avalos](https://reader036.vdocuments.net/reader036/viewer/2022062700/552fc174497959413d8ef188/html5/thumbnails/8.jpg)
KDE-based feature selection
KDE inferida em Probability density function (PDF).
h>0: bandwidth
Gaussian Kernel :
![Page 9: VIOLENT VIDEO DETECTION BASED ON MoSIFT FEATURE AND SPARSE CODING Long XuChen GongJie YangQiang WuLixiu Yao Aluno: Rómulo Ramos Avalos](https://reader036.vdocuments.net/reader036/viewer/2022062700/552fc174497959413d8ef188/html5/thumbnails/9.jpg)
KDE-based feature selection
![Page 10: VIOLENT VIDEO DETECTION BASED ON MoSIFT FEATURE AND SPARSE CODING Long XuChen GongJie YangQiang WuLixiu Yao Aluno: Rómulo Ramos Avalos](https://reader036.vdocuments.net/reader036/viewer/2022062700/552fc174497959413d8ef188/html5/thumbnails/10.jpg)
Sparce coding scheme Ele é mais preciso.
A reduzida a partir do procedimento anterior, passa através de um vector de fórmula discriminativo com a qual se torna um vector de Sparce Code.
Este procedimento é o lugar onde um dicionário que representa os padrões básicos de características de distribuição de dados.
![Page 11: VIOLENT VIDEO DETECTION BASED ON MoSIFT FEATURE AND SPARSE CODING Long XuChen GongJie YangQiang WuLixiu Yao Aluno: Rómulo Ramos Avalos](https://reader036.vdocuments.net/reader036/viewer/2022062700/552fc174497959413d8ef188/html5/thumbnails/11.jpg)
Max Pooling Over Motion FeatureÉ aplicado após a obtenção do conjunto de recursos no Sparce Code.
Elemento pertencente ao vetor de K dimensões:
Zij : Elementos da matriz dada por Sparce Coding
![Page 12: VIOLENT VIDEO DETECTION BASED ON MoSIFT FEATURE AND SPARSE CODING Long XuChen GongJie YangQiang WuLixiu Yao Aluno: Rómulo Ramos Avalos](https://reader036.vdocuments.net/reader036/viewer/2022062700/552fc174497959413d8ef188/html5/thumbnails/12.jpg)
Table of shows: Hockey Fight dataset
![Page 13: VIOLENT VIDEO DETECTION BASED ON MoSIFT FEATURE AND SPARSE CODING Long XuChen GongJie YangQiang WuLixiu Yao Aluno: Rómulo Ramos Avalos](https://reader036.vdocuments.net/reader036/viewer/2022062700/552fc174497959413d8ef188/html5/thumbnails/13.jpg)
Table of shows: Crowd Violence dataset