Pedestrian detection in video of outdoor condition

Pedestrian detection is an important research field in computer vision and a lot of studies have been made to enhance the accuracy rate. However, there are some limitations such as most researchers used to study pedestrian detection with analyzing individual images, instead of considering the relationship of frames in videos. For dealing with pedestrian detection in the video, we present a new method to makes good use of both the image information in a single frame and the difference among contiguous frames. Our approach could improve the detection rate and reduce the false-positive ratio simultaneously, with reaching a speed of processing more than 60 frames per second. Moreover, as for the videos with bad quality and noise, our method also shows performance good enough for real surveillance cases. This paper also built a big dataset including 2414 videos, on which we test our method and achieve good performance. It provides a way to study more in pedestrian detection in videos for other researchers.