计算机工程

• •    

面向嵌入式的高实时微小目标跟踪检测方法

  

  • 发布日期:2020-12-29

Embedded-oriented High Real-time Tiny Target Tracking and Detection Method

  • Published:2020-12-29

摘要: 主流目标检测算法在嵌入式平台无法兼具高实时性与高准确性,进而难以应用于边缘智能等领域,为解决微小目标 跟踪检测在嵌入式平台实时应用的瓶颈,提出一种面向嵌入式设备的高实时微小目标跟踪检测方法,利用轻量化的神经网络 骨干网络和路径聚合网络,并对整体网络进行针对化的剪枝优化,同时深度融合相关滤波算法,提升针对微小目标跟踪检测 的准确度和速度。在 3D 物体场景渲染器自建的军事微小目标数据集上的实验结果表明,该算法在 100 像素的极小目标跟踪 识别中,相比 Darknet53-CSP 等方法检测精度大幅提高;在 400-10000 像素的微小目标识别跟踪中,检测精度和检测速度都优 于 DarkNet53 和 ResNeXt50+CSP 等算法。

Abstract: Mainstream target detection algorithms cannot have both high real-time performance and high accuracy on embedded platforms, therefore, it is difficult to apply to fields such as edge intelligence. In order to solve the real-time application bottleneck of tiny target tracking and detection on embedded platforms, a high-real-time tiny target tracking and detection method for embedded devices is proposed. This research uses the lightweight neural network backbone network and path aggregation network to optimize the overall network for targeted pruning. At the same time, it deeply integrates correlation filtering algorithms to improve the accuracy and speed of tracking and detecting tiny targets. Experimental results on the military small target data set built by the 3D object scene renderer show that the algorithm has greatly improved detection accuracy compared with Darknet53-CSP in the tracking and recognition of 100-pixel very tiny targets; In the 400-10000 pixel tiny target recognition and tracking, the detection accuracy and detection speed are better than those of DarkNet53 and ResNeXt50+CSP.