An optimized mask-guided mobile pedestrian detection network with millisecond scale

With the tremendous advance of deep learning, pedestrian detection has gradually become the basis of various subsequent advanced applications, such as segmentation, tracking, or high-level object-based reasoning. However, There are few mobile-based video detection networks that can be implemented in practical production applications. Therefore, we proposed a super lightweight network based on video, which made the following two major contributions: (1) For each frame, the detection boxes of the previous frame and the RGB channels of the current frame are jointly used as the input of the network, which can improve the detection accuracy effectively with little increase in the network complexity, at the same time avoiding the need to adopt relatively limited video datasets during the training procedure. (2) In order to make full use of the time-related information of adjacent frames, we adopt the method of birth and death which is commonly used in tracking and post-matching filtering method to reduce the jitter of the detection boxes between adjacent frames and improve the overall visual effect. The final result outperforms other tiny networks at nearly the speed of 5ms, by achieving 45.5% mAP on person category of COCO-2017 dataset on MI 8 GPU with at least 90% in detection accuracy.