In optoelectronic surveillance systems, the pixel base adaptive segmenter (PBAS) algorithm, which is widely used in moving objects segmentation, is hard to meet the requirements of real-time applications due to its calculating complication and a large amount of computing parameters. With its pixel-level parallelism, deploying PBAS on top of graphic processing unit (GPU) is promising. This paper implements real-time optimization of PBAS on embedded GPU platform-Jetson TX2, employing methods of data storage architecture, shared memory utilization and random number generation. Experimental results show that the parallel optimization method can achieve 132 fps when processing 480×320 pixel medium-wave infrared video sequences, thus meets the real-time processing need.