Currently, reliable detection and accurate counting of firearm projectiles is one of the difficult points of gun and ammunition management. To improve the accuracy and reliability of projectile detection algorithm based on acceleration signals, we propose a new time-domain feature extraction method for firearm firing signals: The time-domain segmental feature extraction method, which avoids the problem that time-domain features are overly dependent on acceleration transient spikes. Firstly, various statistical features of the sample signals of gunshot acceleration in the time and frequency domains have been extracted. Then machine learning classification algorithms K-nearest neighbors, logistic regression, support vector machines, decision trees and random forests are used for gunshot recognition modeling. Finally, the effects of various single features on the performance of gunshot recognition models are explored and compared. The experimental results show that the extracted main fluctuation domain area feature have the optimal discrimination and can achieve more than 99% classification accuracy on most machine learning algorithms.