Early Detection of Activity in Untrimmed Videos Using Dense Optical Flow and Deep Neural Network

Computer Vision is playing aremarkable role right from essentials to entertainment and thus trying to turn computer as a “seeing” machine. Having widespread applications in most of the real world domain like healthcare, surveillance., space and social media, yet it is still challenging to match human intelligence in understanding the underlying context. Recognition of a task performed by human is important to learn the happenings, but predicting it, from the activities” initial subset becomes an aid to overcome adversities in near future. Numerous algorithms and techniques are present, yet a gap exists due to the environment in which it is recorded, illumination condition,overlapping of men and materials on captured scenes etc.These lead to the requirement of high power computing devices for recognition, which is considerably costly.In this work, Dense OPnet (Dense optical flow and Convolution neural Net) is proposed which takes in the RGB input video, converted to dense optical flow frames and then fed to three layered CNN for recognition and early detection of an activity using N/2 frames of the total input video. This method is evaluated on two benchmark datasetlike J-HMDB21 and UCF50 giving an accuracy of 93.4% and 74.3% on recognition, 68.2% and 50.24% on early action detection respectively.