| Titre : | ABNORMALITY DETECTION THROUGH VIDEO ANALYTICS | | Type de document : | projet fin études | | Auteurs : | Sara BOUHALI, Auteur | | Langues : | Français (fre) | | Catégories : | e-Management et Business Intelligence
| | Mots-clés : | Artificial Intelligence, Deep Learning, Data Mining, abnormality, surveillance cameras, MIL, regression model, weakly labeled data. | | Index. décimale : | 2109/19 | | Résumé : | Surveillance cameras are widely spread all over the place serving principally to monitor multiple scenes for security purposes. However, it is challenging for a human to keep in track with all the tapes in real time. If only abnormal events are automatically identified, we can not only save a lot of human labor but also enhance significantly the surveillance systems. Thus, identifying abnormal events within a livestream of video surveillance is tedious, complex, and a crucial issue.
This work is about a basic approach for real-world abnormality detection by exploiting videos that are depicting both normal and abnormal activities. An attempt is made to propose an intelligent system for surveillance. The model we propose is based on multiple instance learning (MIL) due to the laborious task of specifying the exact temporal annotation of the anomalous segment in each video. In other words, we will use weakly labeled videos for the training of the model since the two labels we have (normal and anomalous) are available only on video level instead of segment level which motivates the intuition behind using MIL.
We consider a video as a bag of a fixed number of segments that we will refer to as instances in MIL. Hence, we draw the assumption that if a video is abnormal then the bag is positive and contains at least one abnormal segment depicting the abnormality, on the other hand we refer to a normal video as a negative bag since all its segments are normal. Our MIL regression model will assign to each segment of a video a score, given that the highest scores are those assigned to anomalous video segments.
We found that using both normal and abnormal videos to train our model gave better results in comparison to previous work on abnormality detection.
Through this work, we first present all the steps required to ensure the management of a machine learning project. Then, we report on the implementation details of our approach and how we evaluate its performance using the AUC/ ROC curve.
|
ABNORMALITY DETECTION THROUGH VIDEO ANALYTICS [projet fin études] / Sara BOUHALI, Auteur . - [s.d.]. Langues : Français ( fre) | Catégories : | e-Management et Business Intelligence
| | Mots-clés : | Artificial Intelligence, Deep Learning, Data Mining, abnormality, surveillance cameras, MIL, regression model, weakly labeled data. | | Index. décimale : | 2109/19 | | Résumé : | Surveillance cameras are widely spread all over the place serving principally to monitor multiple scenes for security purposes. However, it is challenging for a human to keep in track with all the tapes in real time. If only abnormal events are automatically identified, we can not only save a lot of human labor but also enhance significantly the surveillance systems. Thus, identifying abnormal events within a livestream of video surveillance is tedious, complex, and a crucial issue.
This work is about a basic approach for real-world abnormality detection by exploiting videos that are depicting both normal and abnormal activities. An attempt is made to propose an intelligent system for surveillance. The model we propose is based on multiple instance learning (MIL) due to the laborious task of specifying the exact temporal annotation of the anomalous segment in each video. In other words, we will use weakly labeled videos for the training of the model since the two labels we have (normal and anomalous) are available only on video level instead of segment level which motivates the intuition behind using MIL.
We consider a video as a bag of a fixed number of segments that we will refer to as instances in MIL. Hence, we draw the assumption that if a video is abnormal then the bag is positive and contains at least one abnormal segment depicting the abnormality, on the other hand we refer to a normal video as a negative bag since all its segments are normal. Our MIL regression model will assign to each segment of a video a score, given that the highest scores are those assigned to anomalous video segments.
We found that using both normal and abnormal videos to train our model gave better results in comparison to previous work on abnormality detection.
Through this work, we first present all the steps required to ensure the management of a machine learning project. Then, we report on the implementation details of our approach and how we evaluate its performance using the AUC/ ROC curve.
|
|
Faire une suggestion Affiner la recherche
ABNORMALITY DETECTION THROUGH VIDEO ANALYTICS / Sara BOUHALI
