Harmful recognition is a process of recognizing harmful activities in public, private places through video monitoring. Harmful recognition can be used to monitor harmful act in public places through video monitoring. In this paper, convolutional neural network(CNN) algorithm is used with Mobilenet v2 architecture and OpenCV for harmful act recognition. In this model/system, we use various real lifebased videos to train our model/system in which half of the data-set contains harmful videos and another half contains safe videos. The proposed model fulfill the accuracy of 90%.