基于改进YOLOv4的行人检测算法

针对YOLOv4算法在行人检测中精度低，实时性差的问题，提出一种基于YOLOv4的改进算法。首先将MobileNetv2作为主干网络，在减少参数量的同时保证其特征提取能力，同时在MobileNetv2中加入Bottom-up连接，减少浅层信息的丢失；然后在特征融合网络嵌入CBAM注意力机制，增强特征的表现力；最后在分类与回归网络中加入Inception结构，进一步提高检测速度和增加网络复杂度。实验表明：在VOC数据集上，改进算法比原算法检测效果更佳，实时性更好，其精度提高了2.87%，处理速度提升了29.52FPS；同时在真实场景下构建的数据集上，改进后的算法比YOLOv4精度提高了2.13%，具有较好的鲁棒性。

Pedestrian Detection Algorithm Based on Improved YOLOv4

Aiming at the problem of low accuracy and poor real-time performance of YOLOv4 algorithm in pedestrian detection, an improved algorithm based on YOLOv4 was proposed. First, Mobilenetv2 was used as the backbone network to reduce the amount of parameters and ensure its feature extraction capabilities. At the same time, a bottom-up connection was added to Mobilenetv2 to reduce the loss of shallow information; Then, CBAM attention mechanism was embedded in the feature fusion network to enhance the ability of feature expression; Finally, the Inception structure was added to the classification and regression network to further improve the detection speed and increase the network complexity. Experiments showed that on the VOC data set, the improved algorithm had a better detection effect and better real-time performance than the original algorithm. Its accuracy was increased by 2.87%, and the processing speed was increased by 29.52FPS; at the same time, it was improved on the data set constructed under the real scene. Compared with YOLOv4, the accuracy of the latter algorithm was improved by 2.13%, and it had better robustness.