Abstract:

Patrol duty is a security and stability maintenance operation of great significance， but the patrol environment is complex， the object are diverse， and the problem of difficult detection is very prominent， so how to accurately and real-time detect patrol duty objects is of great practical significance.In order to improve the accuracy and real-time detection of patrol duty objects， the YOLOv5 algorithm is improved. In order to suppress the interference caused by the patrol environment， the ECA-Net attention mechanism is combined to improve the saliency of the detected object； and the introduction of BiFPN structure ensures better real-time performance and multi-scale object detection capabilities of the algorithm.Comparing the improved algorithm with the original algorithm， the mAP is improved by 3.51% ； comparing with four algorithms， the results show that the algorithm can better reduce the impact of patrol object detection of due to the problems of similar detection，diverse scales and light interference， which further verifies the effectiveness of the proposed algorithm in the task of patrol duty object detection.

Key words: object detection, patrol duty, attention mechanism, BiFPN（bi-directional feature pyramid network）, deep learning, computer vision

摘要：

巡逻执勤是具有重要意义的安全维稳行动，但是巡逻环境复杂、目标多样、检测难度大的问题十分突出，所以如何准确、实时检测巡逻执勤目标具有重大现实意义。为了提升对巡逻执勤目标检测的准确性和实时性，基于YOLOv5算法进行改进。为抑制巡逻环境带来的干扰，结合ECA-Net注意力机制进行改进，提高被检测目标显著性；同时为保证较好的实时性及多尺度目标检测能力，引入BiFPN网络结构。将改进算法与原始算法进行比较，mAP提升3.51%；与4种算法进行了对比实验，结果显示该算法能较好地降低巡逻执勤目标检测因检测相似、尺度多样、光照干扰等问题带来的影响，进一步验证了该算法在巡逻执勤目标检测任务中的有效性。

关键词: 目标检测, 巡逻执勤, 注意力机制, BiFPN, 深度学习, 计算机视觉