First of all, in order to solve the problem of drift and sensor failure in the process of installation and wearing of a single sensor, an array sensor acquisition system is designed in this paper. Then according to the common five-point array layout, theoretical analysis and comparative experiments are carried out to determine the best sensor layout of the system. Next, a gait dataset with 40 people and 7 patterns is constructed. In view of the problems of global information loss, large computation and memory consumption, and insufficient boundary information processing in the gait recognition network of embedded deployment, improvements are made. An encoder-decoder based parallel attention convolutional network is proposed. Finally, a multi-mode motion gait recognition experiment is set up to verify the performance of the algorithm. The experimental results show that the algorithm can quickly and accurately identify 7 common human gait patterns with an average accuracy of more than 95%, which has good performance.