Unconstrained Face Detection Based on Cascaded Convolutional Neural Networks in Surveillance Video

With the popularity of surveillance video, face detection in surveillance video has become a popular and important topic. Face detection in surveillance video plays an important role in many popular applications such as: personal identification, crowd analysis, database establishment, and abnormal event detection. This paper proposes an unconstrained face detection method for surveillance video, which is not influenced by factors such as face location, expression, posture, scale, and lighting conditions. First, the detection area is initially extracted from the video frame using the improved foreground extraction and skin color detection. Next, we then use the multi-scale sliding window and the cascaded Convolutional Neural Network (CNN) designed in this paper to detect faces. This cascaded network consists of two CNN networks: the first network filters out most of the background area while ensuring the running speed of the whole system and the recall rate of the face, while the second network guarantees the accuracy of the overall system. Finally, we set up a database for the experiment which contained samples from the actual surveillance video. The results of our experiment suggest that the proposed method can obtain good results on unconstrained face detection in surveillance video and can also achieve satisfactory detection speed.