Pedestrian Detection From Video Sequences

Pedestrian detection attracts more and more attention in computer vision area in recent years. It involves many academic disciplines, such as pattern recognition, image processing, and artificial intelligence. It also has important applications in intelligent visual surveillance, assistant driving system, safety control and so on. But pedestrian detection is still a difficult task, because of its non-rigid character, inner-occlusion and between-occlusion, various clothes and attachment, lighting etc.This paper studies the pedestrian detection technology independent of pre-segment in video sequences. It mainly concentrates on how to design effective features to describe pedestrian model. The research results are in the following: This paper proposes a new feature to describe pedestrian model– 3D Haar-like features. Its main idea is to scan densely the Space-time Volume in video sequences using seven types of still and dynamic 3D Haar-like filters. The response of filters forms the representation of the pedestrian. The advantage of this feature lies in that it not only describes the appearance information, but also captures the motion pattern. The experiments show that 3D Haar-like features could effectively model the pedestrian.Considering the influence of background, clothes and attachment and varying illumination, this paper proposes the gradient-based 3D Haar-like features. This method first does a preprocessing stage of Distance Transform to the segmented Space-time Volume. Then it extracts the 3D Haar-like features in the silhouette images. Its advantage is that it is insensitive to noises and reserves the silhouette information at the same time. The experiments show that the gradient-based 3D Haar-like features have better detection performance.This paper proposes a feature Edge Direction Histogram (EDH) feature, considering the effectiveness of gradient features to model the pedestrian. First it extracts the edge to every frame using Canny operator, which could obtain the silhouette and filter the noises. Then it divides the space-time volume into several sub-blocks. In each sub-block, it computes the histogram for each gradient direction of the edge. All the local histograms are concatenated into one feature vector to constrain the global shape and represent the space-time volume. The experiments show that this feature achieves good detection performance, as well as robustness.The above experimental results indicate that 3D information can well describe the appearance and motion information of pedestrian, and achieve good detection performance. Considering the complexity of pedestrian, gradient-based features are more robust than intensity-based features. We believe better detection performance would be obtained by combining the features with better discriminative analysis technique and statistical leaning methods.