Pyramidal-Relative Entropy Based Temporal Signature for Video Transition Detection using LSTM

Abstract Video transition detection (VTD) is a significant topic in the field of video analytics, owing to its useful applications in video indexing, video surveillance, and video understanding. The challenge in VTD is to extract the complex temporal variations caused by the change in illumination, rapid motion of object and camera. To address these challenges, a pyramidal-relative en-tropy based long-short term memory (LSTM) framework is proposed. Initially, the uncompressed video frame is modeled using pyramidal attributes. Then, the temporal signature is generated based on forward-backward ratio of relative entropy measure. The LSTM network is trained to detect the transitions through temporal signature and categorizes them as no transition, abrupt transition and gradual transition. Benchmarks namely TRECVID 2001, TRECVID 2007, VIDEO-SEG2004, RAI and BBC data-sets have been used in evaluation assessments. The simulation results illustrate the efficacy of the proposed framework to achieve an average F 1 score of 95.5±0.06%, 98.51±0.79%, 97.5 ±2.6%, 93.54±5.03%, and 97.58±1.89% on TRECVID 2001, TRECVID 2007 VIDEOSEG 2004, RAI and BBC data-sets, respectively.