Deep learning model training method and system for indoor multi-person identification and detection

The embodiment of the invention provides a training method of a deep learning model for indoor multi-person identification and detection. The method comprises the following steps: carrying out indoor volume activity detection, and determining an indoor noise training set which does not reach a preset power density threshold value and a footstep sound training set which reaches the preset power density threshold value; determining an indoor environment noise signal-to-noise ratio by using the indoor noise training set, and updating an indoor noise library based on the signal-to-noise ratio; filtering additional noise in the footstep sound training set based on an indoor noise library to obtain a first sample feature; determining a probability distribution function of the footstep sound training set through dynamic time warping, establishing a similar matrix by using the probability distribution function, and matching to obtain a second sample feature; and training the deep learning model until the model conver