This video illustrates some of the work I have done with activity recognition. This dataset was collected as part of the CMU Multi-Modal Activity Database, and the modalities include image data from multiple perspectives, internal measurement units (IMUs), infrared motion capture, and wearable accelerometers. 

Our approach leveraged a variant of KNN-SVM, coupled with dynamic temporal smoothing. While recognizing approximately 35 different activities, we were able to achieve 75% classification accuracy. The predictions of our approach are shown in the video on this page. This work was conducted with Prashant Reddy as part of a graduate course project.