Human activity analysis and classification using RGB-D videos

This thesis encompasses parts of the field of computer vision. The main problem dealt with throughout this project is how to automate classification of specific human activities via video streams. The essence of this project is therefore that the developed algorithm shall be able to distinguish these activities from one another. A number of restrictions were imposed on the data-set in order to keep the problem size manageable. A video snippet was assumed to contain a single human that either falls to the ground or, in a controlled way, sits and then lies down on the ground. Our approach relies heavily on background subtraction as the method for detecting foreground objects and comparing the performance of several such approaches for video pre-processing. Features are designed with a heuristic approach utilizing velocities, limb distance relationships and acknowledged techniques based on oriented gradients and optical flow. The classification model is a pre-implemented support vector machine from the libSVM package that is tuned to fit the generated features. The accuracy of our algorithm reached 93.73% which was well above the goal of 85%. On average the classification process roughly took 0.6 second per frame which concludes that the run-time of our algorithm was not fast enough to process a video stream in real time. The result of this work confirms, however, that the methods used works well in a restricted setting and produces a high classification rate. The gist is thus that the proposed method is a good classifier when used in a restricted, offline environment.