An Empirical Study of Block Matching Techniques for the Detection of Moving Objects

Love, N S ; Kamath, C

Lawrence Livermore National Laboratory

关键词: Compression;    Cameras;    99 General And Miscellaneous//Mathematics, Computing, And Information Science;    Detection;    Algorithms;   

DOI  :  10.2172/898460
RP-ID  :  UCRL-TR-218038
RP-ID  :  W-7405-ENG-48
RP-ID  :  898460

美国|英语

来源: UNT Digital Library

【 摘 要 】

The basis of surveillance, event detection, and tracking applications is the detection of moving objects in complex scenes. Complex scenes are difficult to analyze because of camera noise and lighting conditions. Currently, moving objects are detected primarily using background subtraction. We analyze block matching as an alternative for detecting moving objects. Block matching has been extensively utilized in compression algorithms for motion estimation. Besides detection of moving objects, block matching also provides motion vectors (location of motion) which can aide in tracking objects. Block matching techniques consist of three main components: block determination, search methods, and matching criteria. We compare various options for each of the components with moving object detection as the performance goal. Publicly available sequences of several different traffic and weather conditions are used to evaluate the techniques. A coherence metric and the average magnitude of object motion vector error are used to evaluate block determination approaches and search methods. To compare the matching criteria we use precision-recall curves to evaluate the performance of motion detection. We present an empirical study of the block matching techniques using these metrics of performance as well as process timing. We found the hierarchical block determination approach has an overall higher coherence of object motion vectors than the simple block determination approach, but with a significant increase in process timing. The average magnitude of object motion vector for the search methods evaluated were comparable, with the cross search method having a better coherence of object motion vectors. Overall the three step search (TSS) detects more moving objects than the cross and 2D-logarithmic search methods. And the mean square difference (MSD) matching criterion has the best precision-recall as well as process timing when using zero motion biasing.