Research And Implementation Of Moving Target Detection Based On TMS320DM6467

Moving target detection is key technology to intelligent video surveillance, because it isn't possible to be tracked and analyzed for the target which isn't to be detected correctly in the video scene, how to detect the moving target accurately and efficiently in complex video scene is important and difficult point for intelligent video surveillance system. With the development of embedded technology, intelligent video surveillance based on embedded system has been more favored for its stability and convenience. The purpose of this paper is to achieve a product which has practical application in intelligent video surveillance. The contents in this paper include the software design based on TMS320DM6467 platform and the implementation of moving target detection algorithm.TMS320DM6467 is the most advanced media processor in DaVinci technology, it has the dual-core architecture integrated the C64x+ DSP core and ARM926EJ-S core, the DSP core is for calculate and the ARM core is for control. Embedded Linux OS is migrated to ARM core for better system control, so the software design is application development under embedded Linux. The software function include communication with PC and network camera via Ethernet, and C64x+ management by adding a kernel module to Embedded Linux OS, this make the software control the moving target detection algorithm which is executed in C64x+ core. The software design is divided into system control and algorithm implementation. Using this ideal, not only make it convenient for system control achieved by ARM core and algorithm achieved by DSP core, but also make the system maintenance and algorithm improve independence.For moving target detection, this paper is proposed an improved Gaussian mixture model. The proposed algorithm uses the learning rate as the updating step of model's mean and variance, in this way, the background model update speed has been improved and the model can adapt to the video scene faster. Meanwhile, this make it possible to uses the C64x+ special instruction set saving the calculate time in order to ensure the detection in real-time, hi Gaussian mixture model mach step, a dual threshold algorithm is proposed to improve the accuracy of moving target detection. This algorithm can better suppress local noise by using the neighborhood information. Finally, a novel shadow suppression method is used in the post-processing of moving target detection, the post-processing section also includes mathematical morphology filter and connected domain detection for isolated noise points removal and filling the target internal gaps, and the false target can be deleted by using the result of connected domain detection and priori knowledge of targets.