A efficient method for embedded target detection

Abstract:
In order to improve the performance of target detection in embedded devices, an improved YOLOV3-Tiny is proposed in this paper. The detection accuracy is improved by redesigning new backbone, introducing the Spatial Pyramid Pooling (SPP) module, and using new loss function, which is constructed by combining the Giou loss and Focal loss. In addition, a new target dataset containing 2400 images in seven categories is established for experimental verification. In the training phase, the strategy of migration learning is adopted to deal with the problem of insufficient image data. The experimental results show that our method has a 5.5% improvement in mAP@0.5 than YOLOV3-Tiny. At the same time, the multi-threading is introduced for speeding up the model inference, and the measured speed reaches 41FPS, which is slightly higher than YOLOV3-Tiny.