A Research On Stochastic Computing Based Moving Object Detection And Tracking Algorithms

Object tracking and detection is a key technology in many fields such as computer vision and pattern recognition.It has a wide range of applications,including intelligent monitoring,intelligent transportation,human-computer interaction and so on.With the explosive growth of intelligent devices,the importance of embedded systems is continuously growing.Therefore,designing a high performance object tracking detection system with low power consumption and low complexity is of great importance.On the other hand,the stochastic computing is a novel technique for digital signal processing,which uses the random bit-stream to represent and calculate the numbers.In that way,simple circuits can be used to implement the complex operations.This thesis focuses on the designing of the high-performance,power-efficient object tracking and detection system with stochastic computing.Firstly,this thesis introduces the backgrounds of object tracking and detection algorithms and stochastic computing,analyzes the applications of stochastic computing,and demonstrates the rationality of stochastic computing techniques to design the object tracking and detection systems.Through the comparison of the classical object detection algorithms,including background subtraction,frame difference,optical flow and three-dimensional Markov field method,we analyze the advantages,disadvantages and applications of these algorithms.Then,a fast time-space three-dimensional belief propagation algorithm based on the traditional method and a stochastic belief propagation algorithm are proposed to design the object detection systems,respectively.This thesis also proposes a stochastic normalization method,a stochastic matrix multiplication method,and the method of using the positive/negative likelihood function to suppress the noise.The proposed stochastic object detection system can achieve hardware efficiency 85% higher than the traditional realization without performance loss.Finally,this thesis uses the stochastic computing techniques to design the object tracking system with particle filter.The proposed object tracking used the techniques of stochastic exponentiation,two-segmented stochastic multiplication and stochastic matrix multiplication.Compared with the traditional implementation,the hardware efficiency of the stochastic object tracker can be improved by 294% without loss of performance.