Real-Time Activity Detection of Human Movement in Videos via Smartphone Based on Synthetic Training Data

Current research in the domain of video activity detection focuses on real-time activity detection. This includes multiple approaches in the mobile environment, such as the detection of correct motion sequences in the sports and health area or in safety-relevant environments. Current trends focus on the use of 3D CNNs. This work describes a approach to combine the results of a human skeleton point detector with an LSTM on mobile devices. Frameworks for pose detection are combined with LSTMs for activity detection with sensor data, optimized for the mobile area. The resulting system allows the direct detection of a person pose and the classification of activities in a video recorded with a smartphone. The successful application of the system in several field tests shows that the described approach works in principle and can be transferred to a resource-limited mobile environment by optimization.