Jerky Action

Fast-motion playback of the movie, for example, with the help of the fps (frames per second) parameter in MPlayer, would be the easiest option:

mplayer -framedrop -fps 150 video.mp4

The -framedrop parameter simply throws away frames if the CPU fails to keep pace when decoding. The result is a movie that runs roughly five times as fast as normal, which looks like something from the early days of movie making, like the old hand-cranked takes of Charlie Chaplin.

This procedure requires a human reviewer, whereas I was looking for an automated process that lists the most important seconds of the retrieved video as metadata, along with illustrative thumbnails, much like a photographer's contact sheet.

Motion Detection

Much has happened in the field of pattern recognition in image processing in the last few years, and the OpenCV [2] package even offers some highly scientific routines as an open source program. To determine whether someone or something is moving through the video, a program needs to read the frames in the video stream and determine which pixels have shifted from the (x, y) coordinates to the (x + Δx, y + Δy) coordinates. If you find a large contiguous area for which this is true, then you can assume that a movement took place between two frames.

One of the procedures included with OpenCV is called Lucas-Kanade [3]; it attempts to find optical flow – areas around central points that jointly move from one frame to the next – in a video. To do so, it first determines a number of interesting areas, where monitoring promises success and that another algorithm can extract from an image by focusing on image points in areas with a recognizable structure or on the edges of objects.

To perform the Lucas-Kanade (LK) analysis, the OpenCV package (on Ubuntu, this is libopencv-dev) provides the calcOpticalFlowPyrLK() function with no fewer than 11 parameters.

Acrobatics with Cmake

The C++ program in Listing 1 [4] reads in a video file and detects any movement of objects between frames. Converting it into an executable program requires some compilation acrobatics with include files and link libraries; your easiest approach here is to use cmake and its meta Makefile in Listing 2.

Typing the cmake . command (the dot stands for the current directory, in which the CMakeLists.txt file resides) followed by the make command starts the lengthy compilation process, which finally produces the max-movement-lk binary. It expects a video file and outputs movie location values in seconds for scenes containing motion.

To do this, the main() function reads the video file name from the command line and starts VideoCapture provided by the OpenCV package in line 57. The frame rate is read from the video file in line 63 and stored in the fps variable. Because the LK algorithm works best with grayscale images, lines 68 and 75 bleed the color out of every frame to be analyzed.

A while loop iterates across all the frames, and the move_test() function checks in line 76 whether any motion occurred between the last frame read, oframe, and the current frame. If so, line 77 divides the counter value by the FPS value of the video and thus computes the time at which the motion occurred in the video as a value in seconds.