Towards Real-time Object Detection on Edge Devices for Vehicle and Pedestrian Interaction Scenarios

Object detection in complex road environments has always been an important and challenging task in assisted driving systems for autonomous driving. However, the low accuracy or the inability to achieve real-time performance limits the application of current object detectors in autonomous driving. In view of the difficulty in real-time and accurate detection of vehicles and pedestrians in a road environment, a lightweight detection network (MRS-YOLOv3) applied on edge devices is proposed based on the structure of YOLOv3. Combining the multi-receptive field spatial pyramid pooling block and the bidirectional feature pyramid path aggregation structure, the output feature map of the backbone is interacted in the spatial domain and the scale domain. By introducing not-adjacent scale feature interaction module before multi-scale feature aggregation, cross-scale features can be efficiently interacted. In addition, we also use DIoU Loss and Focal Loss as the loss function to make the model achieve better performance. Finally, we deployed the proposed model to the edge device Jetson TX2 for actual evaluation. The results show that MRS-YOLOv3 can perform real-time and efficient detection in vehicle and pedestrian interaction scenarios, achieving a better trade-off between detection accuracy and speed.