An Iterative Learning Scheme with Binary Classifier for Improved Event Detection in Surveillance Video

Abstract

This paper presents an iterative training framework with a binary classifier to improve the learning capability of a deep learning model for detecting abnormal behaviors in surveillance video. When a deep learning model trained on data from one surveillance video is deployed to monitor another video stream, its abnormal behavior detection performance often decreases significantly. To ensure the desired performance in new environments, the deep learning model needs to be retrained with additional training data from the new video stream. Iterative training requires manual annotation of the additional training data during the fine-tuning process, which is a tedious and error-prone task. To address this issue, this paper proposes a binary classifier to automatically label false positive data without human intervention. The binary classifier is trained on bounding boxes extracted from the detection model to identify which boxes are true positives or false positives. The proposed learning framework incrementally enhances the performance of the deep learning model for detecting abnormal behaviors in a surveillance video stream through repeated iterative learning cycles. Experimental results demonstrate that the accuracy of the detection model increases from 0.35 (mAP = 0.74) to 0.91 (mAP = 0.99) in just a few iterations.