基于分离卷积的战场目标聚集行为识别

针对大部分行为识别算法效率较低,难以应对大规模影像识别任务的问题,一方面,提出一种结合双流结构与多纤维网络的双流多纤维网络模型,分别以RGB序列、光流序列为输入提取视频的时空信息,然后将两条支路网络的识别结果进行决策相加,提高了对战场目标聚集行为的检测效率与识别准确率;另一方面,提出一种结合分离卷积思想与多纤维网络的双流分离卷积多纤维网络模型,进一步提高网络检测效率与抗过拟合能力。实验表明,在建立的情报影像仿真数据集中,上述算法能够有效识别出战场目标聚集行为,在大幅提升检测效率同时实现了识别准确率的提升。

Battlefield target aggregation behavior recognition based on separation convolution

Due to low efficiency,it is difficult for most behavior recognition algorithms to cope with the identification of large-scale videos.On the one hand,the two-stream multi-fiber network model is proposed by combining two-stream structure and multi-fiber network.It takes the stacked RGB frames and the sequential optical flow images as input to extract spatio-temporal information of video.The results of RGB network and optical flow network are combined with fusion.The two-stream multi-fiber network model improves the detection efficiency and recognition accuracy of battlefield target aggregation behavior.On the other hand,by combining the idea of separation convolution with the multi-fiber module,two-stream separate convolution multi-fiber network model is proposed to further improve the detection efficiency and reduce overfitting.The experiment shows that in the intelligence videos simulation datasets,the above algorithms can effectively identify the battlefield target aggregation behavior and greatly improve the detection efficiency and recognition accuracy.