Moving object detection based on Mixture of Gaussian fusing spatial-temporal information

Abstract

According to the characteristics of embedded system which is severely restricted in computational resource and memory, we propose a novel real-time detection algorithm based on Mixture of Gaussian fusing spatial-temporal information. The algorithm dynamically adjusts the number of Gaussian distributions according to the complexity of background, which can greatly improve the computational efficiency and reduce the memory consumption on the basis of proper detection. In addition, we propose a spatial-temporal histogram algorithm based on block to correct or verify the extracted object pixel, which can reduce the error detection caused by dynamic background. Eventually, we implement our algorithm on multi-DSP system based on TI's TMS320DM642. The experimental results show that the proposed algorithm achieves good performances in real time and proper rate of detection (recall and precision). In conclusion, our method is able to accomplish the detection of moving object in complex scenes stably and reliably.