Research Of Small Moving Targets Detection Based On DSP

Currently in the field of target surveillance,when a sequence of images captured by front-end acquisition system has been received in the afterward processing system,how to make use of these continuous image frames to extract moving targets in the surveillance area is a set of key problems.And when in low SNR scenarios,detecting and tracking small moving targets is especially difficult.If surveillance facilities are capable of recognizing small moving targets in real time application scenarios without human participation,it gives great convenience to other systems or agencies for providing important information about surveillance region in time.This paper based on the theory of Bayes probability framework,using finite set theory and probability hypothesis density filter to achieve the above goal.Probability hypothesis density filter is referred to as PHD filter,adopting the thought of Bayes estimation to estimates targets states.The mathematical form of the likelihood function for multiple targets may be quite complex,therefore,the first moment of the multiple targets posterior probability density function is used to approximate the motion states of the targets.In order to achieve such a complex filter in Engineering,adopting particle filtering method to put PHD filter into practical use.Among the method,every particle represents a state that a specific target may be currently in,and every particle weight represents the probability that the target is in that state.When the number of targets is time varying or background noise is catastrophic,the working status of the PHD filter may be unstable.In order to overcome the problem measurement noise brings in,the filter is improved by introducing the concept of smoothing in mathematics.Smoothing improves the stability of the filter by sacrificing time complexity and computation quantity of algorithm.The simulation results show that the smoothing makes the estimation of targets number more stable.In order to put such complex filter into practical use,this paper turn to multicore DSP for help.TI's multicore DSP TMS320C6678 has eight cores with up to 10 GHz processing power,providing shared memory space,direct memory access between inner components.This paper adopts Inter Processor Communication to achieve multicore scheduling.Particle filtering,the most computational process in the algorithm is decomposed and dispatched to eight cores for parallel processing.Among all the cores,the main core is also responsible for scheduling,dispatching and merging results.According to the analysis of parallel particle filtering,the mathematical model of scheduling process is built.And the scheduling problem is finally converted to two constraint optimization problems.By comparing time consuming between single core and multicore,the real-time performance of the filter has been significantly improved by our multicore DSP implementation.Generally speaking,the processing speed reaches 15 frame/s under measurement resolution of 1024*768.