Event Detection In Surveillance Video

Surveillance video based fall detection is one of the most important task in smart surveillance system. We have a research mainly for event detection in surveillance video, which include three part: pedestrian detection, multiobject tracking and fall detection.With the development of deep learning model, there are more sollutions for problems in computer vision. Faster-RCNN structure is improved by merging the feature maps in VGG-net, which increases the resolution of feature maps and improves the precision of detection. The experiments in public datasets show that the method improves the accuracy rate and recall rate.We accomplish and improve a tracklet confidence based multiobject tracking algorithm. The tracking problem is solved in two stage. Local association is used by associating high confidence tracklets with objects in first stage. Global association is used by associate low confidence fragmented tracklets with tracklets and o'bjects in second stage. We propose a method to eliminate the impact of the size feature caused by perspective and image distortion. This method uses the average width and height in each position to normalize width and height of bounding box.The association is more robust for long time fragmented tracklets by using this method. The experiments show that the improvement is effective.Two kinds of fall detection algorithms are designed and accomplished in this thesis. The first one is based on the traditional method which combines the Cmotion feature sequence, posture feature sequence and Hidden Markov Model to detect fall events. We propose a method to distinguish upright (walking or standing) person from the persons with other postures: projection length of a person in image plane is used as a feature and combine other features to classify the postures. The second fall detection algorithm is based on deep learning. Optical flow and gray image are used to classify the posture of a person and posture sequence is used to detect fall events.