Abstract: Aiming at the problems of low accuracy, large amount of parameters, large amount of calculation, and large model volume, difficulty of being implemented on embedded platforms with limited resources when YOLOv3 algorithm is applied to human body detection, an improved YOLOv3 and its model compression algorithm are proposed. By introducing dense connections and multi-branch structure in YOLOv3, increasing the network width and multi-scale receptive field, and strengthening feature reuse, the accuracy of the model is improved. For the improved YOLOv3, channel pruning is performed by jointly optimizing the weight loss function and the L1 regular term of the BN layer scaling factor, thereby reducing the amount of parameters and calculations, and the volume of the model is greatly compressed. Experimental results show that the accuracy of the improved YOLOv3 algorithm is increased by 6.01%, and the model volume is reduced by 38.46%. After compression, although the accuracy of the model is reduced by 3.16%, the model volume is only 3.31% of the original, only 4.77 MB. Therefore, the improved and compressed YOLOv3 still maintains a high accuracy rate, and the model volume is greatly reduced, which provides support for the YOLOv3 model to realize human body detection in embedded deployment.

Key words: YOLOv3, human body detection, dense connection, multi-branch structure, model compression