Інерційна система розпізнавання жестів

The paper is devoted to the implementation of the gestures recognition system. Existing gesture recognition approaches are considered. Among them the system of inertial type is chosen for research, since the systems of this type are less difficult to implement and require relatively low computational complexity. Thus, the proposed system is based on two components: the hardware module for gesture capturing and the subsystem for the recognition of captured gestures. The hardware component of the system employs MEMS accelerometer and gyroscope for motion acquisition. The prototype of this component is implemented using the STM32F401RE microcontroller and the MEMS sensor MPU9250. The communication between the microcontroller and the MEMS sensor is carried out via the I2C bus. The microcontroller initializes the sensor and acquires the motion data. Then it sends the collected data via the virtual COM port to the PC, where the captured motions are recognized. The system component for the recognition of gestures is implemented as the neural network in Matlab environment. This network consists of four layers: the input layer, the BLSTM layer, the fully-connected layer, and the Softmax layer. The key component of the selected neural network architecture is the BLSTM layer, which (due to its properties) provides some invariance to the length and the shift of the gesture signal in time. The suggested system was tested on three gesture classes: circular motion, tapping and other/unknown gesture. Using the hardware module the training and the testing sets were collected. In course of experiments, it was established that the best training results for the selected network architecture and the collected training set are achieved using the Adam optimizer. The experiment with different number of hidden units in the BLSTM layer showed that the highest recognition rate may be achieved when 100 units are used. Particularly, the suggested system with the specified number of hidden units in BLSTM layer was able to attain the recognition error less than 3.5% (recognition rate of 96.7%) on the test set.