An Object Tracking Algorithm Based on Particle Filter Combine with Diamond Search

• Main Authors: Meng-Ling Song, 宋孟齡
• Other Authors: Jeng-Shyang Pan
• Format: Others
• Language: zh-TW
• Published: 101
• Online Access: http://ndltd.ncl.edu.tw/handle/05016365704021804752

碩士 === 國立高雄應用科技大學 === 電子工程系 === 100 === Object tracking is one of the key technologies in intelligent video surveillance system. Its purpose is quickly and efficiently to extract the object, and provide the object size, position and speed in each frame. This thesis proposed an object tracking algorithm based on particle filter combined with diamond search (DSEPF). As the traditional particle filter is not easy to estimate the best location in the case of fewer number of particles, but too many the number of particles will cause the increase in the computational complexity. Therefore, this thesis use the particle filter estimates its initial position, after that we use diamond search method to estimate a more accurate solution. This thesis using the redistribution threshold in accordance to the number of particles, each times whether to perform Resampling techniques, to solve the degradation phenomena and depletion phenomenon of particle samples. In the experiments, we focused on two different scenarios and different number of particles, by calculating mean square error, non-overlapping domain and mean time to assess. Finally, the results show that our method using only 150 particles for state estimation will be able to reach the root mean square errors values estimated by the adaptive particle filter with 300 particles for state, using fewer number of particles tracking threw able to maintain accuracy and stability.