Collaborative batch learning for crime scene detection

Toh, Yue Xiang (2022) Collaborative batch learning for crime scene detection. Final Year Project, UTAR.

Preview

PDF Download (2975Kb) | Preview

Abstract

Surveillance camera is used in many settings to capture the real-life happenings. Lack of intelligent surveillance camera system decrease the effectiveness of surveillance camera in reducing crime. Our project developed a system to automatically detect crime scene event from the surveillance camera. In our project, we trained our model with normal and crime video from UCF crime dataset. Our work used I3D model pretrained on kinesis dataset to extract the feature frame by frame. We added an 1D dependency capturing attention module on top of the feature extractor to make the features extracted more useful and suitable for the dataset we were using. We used Multiple Instance Learning network as the framework of our system. Since, it was a weakly supervised learning model, the dataset that we used to train our model is weakly labelled dataset, this means that our dataset will not consist of the exact temporal segment where the anomalies happened in the surveillance video. Ranking loss function with sparsity and temporal smoothness constraint was used as our loss function to better detect the anomaly segment throughout the surveillance video.

Actions (login required)