Adaptive step detection method combining CNN-BiLSTM-SA motion pattern recognition

With the popularity of location based services (LBS), smartphone-based pedestrian step detection methods have important impacts on pedestrian dead reckoning (PDR). We propose an adaptive step detection method combining CNN-BiLSTM-SA motion pattern recognition to address the problem that traditional methods have large step counting errors under multiple pedestrian motion patterns. Firstly, the motion patterns are classified according to the walking characteristics of pedestrians, and the local features of different motion patterns of pedestrians are extracted by using convolutional neural network (CNN), and the weights of the extracted motion features are assigned by using self-attention (SA) mechanism, and then the bidirectional long short term memory (BiLSTM) network is combined to mine the pre-post temporal relationship of pedestrian motion features for classification and recognition. Then the peak detection algorithm with two feature constraints, adaptive minimum peak distance and adaptive dynamic threshold, is proposed to detect the step frequency according to the classification results, and the threshold size is dynamically adjusted in walking. The experimental results show that the average error rate of the proposed method for step frequency detection under eight combined motion patterns is 1.31%, which is 5.97% lower than that of the traditional peak detection, and also better than the fixed threshold method.