Evolution of Trajectories
Neelay Pandit, Sherine Abdelhak · 2017
Achieving high classification accuracy remains a challenge for human action recognition approaches that are based on convolutional neural networks (CNN). CNN-based action recognition methods on resource-constrained edge systems are currently unable to train on whole videos due to infeasible computational and memory requirements. On the other hand, approaches that utilize video-level supervision with sparse-sampling designs run the risk of learning local features shared with multiple similar classes. Additionally, features captured by motion estimation algorithms for temporal stream CNN's are already reduced in dimension, increasing the possibility of label mismatch in methods that rely on short-term features. To address the aforementioned points, we design a novel temporal representation to capture a 1) long-term interval of motion and 2) integrate the trajectory of motion captured therein. We compare our approach with current motion representations and demonstrate its efficacy for examples containing local features with high inter-class similarity. We implement our representation as part of two and three-stream CNN's and conduct experiments on one of the popular and most difficult action recognition datasets: HMDB51. Our results show that, for methods employing sparse-sampling designs, our approach surpasses the current state-of-the-art approaches achieving 71.76% on HMDB51.