多标签多目标融合下国产AI芯片的行人目标检测技术

Pedestrian Target Detection Technology Using Domestically Produced AI Chips Under Multi-label and Multi-target Fusion

行人目标检测需要处理大量的图像或视频数据,在实时性和效率方面对AI芯片的计算能力和存储容量提出了较高要求。为此,提出多标签多目标融合下国产化AI芯片行人目标高精度检测技术。构建行人目标特征提取模型,并对特征信息跨维拼接。利用AI芯片的并行计算能力对以上特征融合计算,结合整流函数获取目标空间聚合特征。基于此,考虑多目标多融合,采用IOU阈值计算行人目标候选框的概率,由此实现行人目标的高精度检测。实验结果表明,利用所提方法检测行人目标具有较高的准确度。

Pedestrian target detection needs to process a large amount of image or video data,which puts forward high requirements on the computing power and storage capacity of AI chip in terms of real-time performance and efficiency.Therefore,the localization AI chip pedestrian target detection technology with multi-label and multi-target fusion is proposed.The feature extraction model of pedestrian target is constructed,and the feature information is spliced across dimensions.The parallel computing capability of AI chip is used to fuse the above features.The convergent characteristics of the target space are obtained by combining the rectification function.Based on this,considering multi-target and multi-fusion,the IOU threshold is used to calculate the probability of pedestrian target candidate frame,so as to achieve high-precision pedestrian target detection.The experiment results show that the proposed method can detect pedestrian targets with high accuracy.