Surveillance videos can capture a variety of realistic anomalies/unlawfully activities but there is no real-time analysis of those captured feed. In this project we propose, two different methods to detect anomaly activities in real-time, 1. Multiple classes by detecting single activity in real-time and 2. Normal activities and Unlawful activities. Our proposed algorithm for video time-scale squeezing was able to standardize the UCF-Crime data-set by rewriting each video with a fixed time length of 1 second in total creating a new standardized data-set.The proposed methods were trained on a 3D residual neural network (ResNet 3D 18) with our unique data prepossessing method along with our algorithm for data-set augmentation which achieves significant improvement in anomaly detection performance as compared to the state-of-the-art approaches. Our prediction method itself is 25-30 times faster than any other method/script/algorithm available on the Internet.It is capable of analysing long untrim videos and segmenting the unlawfully/anomaly activities in an efficient way
The original data-set was not standardises to use for deep-learning purpose so in order to make it suitable we propose an algorithm for video time-scale squeezing was able to standardize the UCF-Crime data-set by rewriting each video with a fixed time length of 1 second in total creating a new standardized data-set of 139 MB.
The VTSA algorithm was design to fast-forward a video in order to get the maximum features in less amount of time.The VTSA algorithm works in three phases.
- First phase: Getting the time Getting the time of each video is necessary as each videos has different amount of time,having a dynamic syntax is what is necessary.
- Second Phase: Calculating the speed up value and Speeding up After getting the time of each video we calculate the speed up time by √ time + 6 where we remove the root of time and add with a constant,thereby fast-forwarding/speeding-up the video by that amount of time. 17
- Third Phase: Condition and save After getting the new time line of the video we put a condition in which the video gets a new speedup value which is represented as √ time + 0.7 and gets seed-up until the time is equal to one. Once the condition is satisfied it breaks the loop and saves the video into .mp4 format or once’s the condition reaches a count up-to 30 it breaks the loop and saves the video into .mp4 format
