An Enhanced YOLOv5 Model with Attention Module for Vehicle-Pedestrian Detection

Vehicle and pedestrian detection is one of the vital elements of autonomous driving environmental perception. Although enormous methods have been proposed for detection tasks, the demand for detecting specialized categories is still noteworthy. In this paper, we propose a model which fuses an attention module by optimizing YOLOv5 algorithm and adjusts the sizes of anchor boxes to further improve the performance of vehicle-pedestrian detection. An attention module can effectively collect discriminative features from dense contextual information, which contributes to the better detection of distant objects. An anchor box adjustment makes the model more familiar with the dataset. The qualitative and quantitative experimental results on KITTI benchmark show that our model outperforms the state-of-the-art methods. The preliminary experiment with an embedded device, Nvidia Jetson TX2, further confirms the realtime capability of our proposed model.