Scholarlink

<p>Nowadays, video surveillance plays a vital role in several real-time applications. The goal of Multi-Objective Detection and Tracking (MODT) is to identify and locate multiple objects from video captured by numerous cameras. It has emerged as an essential role in intelligent transportation with the recent technological improvements. As a result, the field of MODT has generated many excellent research outcomes. Existing detection approaches show poor generalization performance in extracting deeper features in complex scenarios such as occlusion, illumination changes, and dense traffic. To overcome these challenges, this method introduces an enhanced variant of YOLOv5 combined with an optimized DL framework to improve the performance of MODT. Initially, the preprocessing is performed by converting the input video into video frames, applying bilateral filtering to remove noise, and performing background subtraction to improve frame quality. Next, multi-object detection (MOD) is accomplished using the Octave convoluted EfficientNetv2 (OEFFNetv2) backbone integrated into the YOLOv5 (OEYOLOv5) model for feature learning. Subsequently, the dwarf mongoose optimized long short-term memory (DMLSTM) is employed for temporal modeling in tracking, and the hyperparameter is fine-tuned via a good point set with Lens theory-based Dwarf Mongoose Optimization (GLDMO). Experimental outcomes on the UA-DETRAC, MOT20, and KITTI datasets demonstrated that the proposed system enhances detection and tracking performance by attaining a maximum mean average precision (mAP) of 98.55%, 98.33%, and 98.74%, outperforming existing YOLOv5 + DeepSORT, YOLOv8 + ByteTrack, FRCNN + FairMOT, and ST + Simple Online and Real-time Tracking (SORT) methods. Furthermore, the framework maintains high performance under challenging conditions, confirming its robustness in real-world scenarios.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Enhanced YOLOV5-LSTM Framework for Multi-Object Detection and Tracking in Surveillance Videos

  • 期刊论文

  • 发表日期：

    2026年1月27日

查看全文
  • G. Priyadharshini,
  • P. Ponni,
  • Sukanya Varshini

摘要

Nowadays, video surveillance plays a vital role in several real-time applications. The goal of Multi-Objective Detection and Tracking (MODT) is to identify and locate multiple objects from video captured by numerous cameras. It has emerged as an essential role in intelligent transportation with the recent technological improvements. As a result, the field of MODT has generated many excellent research outcomes. Existing detection approaches show poor generalization performance in extracting deeper features in complex scenarios such as occlusion, illumination changes, and dense traffic. To overcome these challenges, this method introduces an enhanced variant of YOLOv5 combined with an optimized DL framework to improve the performance of MODT. Initially, the preprocessing is performed by converting the input video into video frames, applying bilateral filtering to remove noise, and performing background subtraction to improve frame quality. Next, multi-object detection (MOD) is accomplished using the Octave convoluted EfficientNetv2 (OEFFNetv2) backbone integrated into the YOLOv5 (OEYOLOv5) model for feature learning. Subsequently, the dwarf mongoose optimized long short-term memory (DMLSTM) is employed for temporal modeling in tracking, and the hyperparameter is fine-tuned via a good point set with Lens theory-based Dwarf Mongoose Optimization (GLDMO). Experimental outcomes on the UA-DETRAC, MOT20, and KITTI datasets demonstrated that the proposed system enhances detection and tracking performance by attaining a maximum mean average precision (mAP) of 98.55%, 98.33%, and 98.74%, outperforming existing YOLOv5 + DeepSORT, YOLOv8 + ByteTrack, FRCNN + FairMOT, and ST + Simple Online and Real-time Tracking (SORT) methods. Furthermore, the framework maintains high performance under challenging conditions, confirming its robustness in real-world scenarios.