Target tracking is an important part of computer vision, specially the pedestrian detection and tracking is a crucial and difficult field. Many researchers have been devoted to the improvement of target detection and tracking methods. With the wide application of deep convolution network, the result of pedestrian detection and tracking has been improved. However, some complex scenarios are difficult to identify and track by present methods. Therefore, it’s necessary to propose an optimal algorithm to improve the performance of pedestrian detection and tracking. The region proposal network, which included multilayer competitive fusion model was used as pretraining network, and longterm and shortterm update strategy in pedestrian tracking task. The pretraining network applied VGG16 to extract feature maps, and then they were put into the multilayer competitive fusion region proposal network to generate more accurate candidate targets. The online pedestrian tracking algorithm was initialized by the pretraining region proposal network, and the region proposal network was finetuned through 500 positive samples and 5000 false positive examples from the first frame, and then created frame index datasets for longterm and shortterm update. Finally, the pedestrian tracking algorithm with continuous updating of region proposal network was accomplished. The model was verified by experiment and in the public datasets named by Caltech, ETH, PETS 2009 and Venice. The test result showed that the region proposal network which included multilayer competitive fusion model had the perfect performance in pedestrian detection and tracking task, and showed good effects in complex background environment.