Query-By-Object Based Video Synopsis

In this paper, we propose a framework for query-by-object(s) based video synopsis. Video Synopsis aims to create a summary of video by retaining the important activities/events present in the input video. We propose creating a shorter video by selecting salient frames based on the important objects present in the video. We train the Yolov3 model with surveillance videos for object detection. Select the frames as salient based on the importance of objects present in a frame and generate the video synopsis with the salient frames. We demonstrate the proposed method on the Summe and TV Sum dataset and own dataset captured from the surveillance camera. We obtain the average F1 score as 93% and average accuracy as 94%. And also show that proposed method gives better results as compared to VASNET model.