Fall direction detection using parallel CNN based on millimeter-wave radar multidimensional feature fusion

Falls pose a serious threat to the safety of the elderly. Accurately detecting falls and their direction is crucial for medical personnel to assess injury locations and promptly formulate treatment plans. Traditional methods primarily integrate distance and Doppler features of fall behavior, neglecting a significant amount of spatial features. Additionally, the determination of fall direction is easily affected by environmental noise and limb-induced Doppler interference. To this end, this paper proposes a parallel convolutional neural network fall direction detection method based on millimeter-wave radar multidimensional feature fusion, which introduces the fall angle as spatial information while effectively attenuating the interference of environmental noise and limb micro-Doppler frequency to improve the accuracy of fall direction detection. Specifically, this method utilizes techniques such as pulse compression, MTI, Range Fast Fourier Transform (RFFT), STFT, and Capon beamforming to obtain the range-time, Doppler-time, and angle-time features from the human motion state to the falling state. Then, the Doppler-time-map (DTM) and angle-time-map (ATM) after the accumulation of multiple frames are subjected to column threshold feature extraction and column normalization feature processing, respectively, to achieve the effect of scrambling frequency removal and feature enhancement. Finally, the processed feature spectrograms are input into the parallel convolutional neural network for independent extraction of each dimension's features and multi-dimensional fusion, enabling the detection of the fall direction. The experimental results show that the proposed method achieves an average recognition accuracy of 97.52% for different fall directions, compared to an improvement of 2.31% over the traditional feature fusion method.