These videos are test results from our method of detecting moving objects in a surveillance video by using compressive sensing. Compressive sensing is a new mathematical tool for signal processing that makes it possible to acquire and represent a signal using far fewer samples than what is required by Nyquist sampling rate.
In our method, a surveillance video is either acquired or transformed into compressive measurements, which is a form of compression to reduce the bandwidth requirement of transmitting the video in the network. After transmitted to a processing center, the measurements are used to reconstruct simultaneously the background and moving objects in the surveillance video. A main advantage of this method is that the background and moving objects can be reconstructed by using an amount of data that is far fewer than the total number of pixels as in the traditional methods.
In our method, the background of surveillance video is model as a low rank matrix, and the moving objects are modeled as a sparse matrix. Therefore, the background and moving objects are reconstructed by using a low rank and sparse decomposition. The decomposition is performed by processing the compressive measurements of the surveillance video.
We apply our method to four test video sequences, and present the results as shown in the video files. For each video sequence, the result is presented by showing the original video, reconstructed background and moving objects together. The original video is shown in the middle of a frame. The reconstructed background is shown on the left and the moving objects are shown to the right.
We also note the number of measurements used in the reconstruction in terms of a percentage as compared to the total number of pixels in a video. 100% means that the number of measurements used is equal to the total number of pixels in the video.
Browse2: 4% measurements used
ShopAssistant1Front: 4% measurements used
Traffic: 6.67% measurements used
Daniel light: 10% measurements used
The results demonstrate that the background and moving objects can be accurately reconstructed with only a fractional data as compared to the total number of pixels in the video.
The details of the method can be found in the paper: