With the video captured in driving, the surroundings can be sensed economically and conveniently by using object detection techniques, but the accuracy and speed of detection requires a lot in such kind of real-time scenes. In this work, a deep learning-based one-stage object detection algorithm called YOLOv3 is studied. Self-attention mechanism is introduced into this method, by embedding Gaussian mask self-attention modules in the high layers of YOLOv3 network. These modules can merge more global information into feature map to improve the accuracy of model. According to the results of experiments, trained on the MS COCO 2017 dataset, the mAP@0.5 and precision of this improved model can reach to 56.88% and 65.31%. Compared with YOLOv3, its mAP@0.5 and precision increase by 2.56% and 3.53%. Although there is a little loss in detection speed, detection accuracy is significantly improved when the method is applied to assisted driving system.