基于DRN和FasterR-CNN融合模型的行为识别算法

针对传统单人行为识别算法易受行人形态多样性、背景和光照等影响的问题进行研究。基于扩张残差网络（DRN）的精准分类效果及目标检测网络Faster R-CNN在目标追踪方面的准确性，提出了一种DRN和Faster R-CNN的融合网络模型。该模型在Faster R-CNN中融入DRN的扩张卷积残差块代替原来的一般卷积层，并对融合模型进行了两方面的改进：在每一层前面添加一个batch normalization层；用三层扩张卷积残差块代替部分两层残差块。实验结果表明三种融合网络识别算法在Olympic sports dataset数据库上较其他行为识别算法取得了更高的mAP。其中，包含三层扩张卷积残差块的融合模型识别性能最好，mAP达到78.9%。

Behavior recognition algorithm based on DRN and Faster R-CNN fusion model

Due to the traditional single person behavior recognition algorithm is easily affected by the diversity, background and illumination of pedestrians based on the accuracy of DRN in classification and detection network Faster R-CNN in target tracking, this paper proposed a fusion network model composed of DRN and Faster R-CNN. The model integrated with dilated convolution residual in Faster R-CNN to replace the original convolution layer. It also made two improvements to the fusion model. It added a batch normalization layer in front of each layer and used three levels of dilated convolution residual blocks to instead of partial two levels of residual blocks. The experimental results show that the three fusion network recognition algorithms proposed in this paper have achieved a higher mAP than other behavior recognition algorithms on the Olympic sports dataset. Among them, the fusion model with three layers of convolution residual blocks has the best recognition performance, and mAP achieves 78.9%.