Lightweight Object Detection Method Based On Cloud Edge Collaboration

Due to the powerful computing power and storage capacity of the cloud computing center,modern intelligent video surveillance systems are basically implemented by using the central cloud architecture.However,with the dramatic increase of network cameras and the continuous improvement of video quality,the video surveillance based on the central cloud architecture faces more and more challenges.Transmitting large amounts of real-time video data not only requires high network bandwidth,but also easily leads to high transmission latency,making it difficult to meet the real-time requirements of surveillance tasks.Cloud-edge collaborationbased architectures extend computing power down to the network edge near sensors,providing promising solutions to address the high bandwidth requirements and latency-sensitive problems in intelligent video surveillance.However,deep learning-based object detection models usually possess large number of parameters,complex structure,and high requirements on computer resources.The edge devices in the cloud-edge collaborative architecture cannot meet their deployment requirements.Therefore,this paper proposes a lightweight object detection method based on cloud-edge collaboration to solve the problems of high volume,unstable transmission and high latency of traditional video surveillance real-time data transmission.The details of the research are as follows:First,significant progress has been made in the generalization capability and accuracy of deep convolutional neural networks,which have a key role in machine vision areas such as image classification and 3D reconstruction,and object detection.However,the computational and storage resources of edge devices are still much lower than those of cloud servers,making it difficult to migrate such models for deployment to resource-limited edge devices.In this paper,we analyze the redundancy in DenseNet dense connections and propose an efficient Single Compression Dense Block(SCDB)based on dense connections.A lightweight DenseNet-based convolutional neural network is constructed with this module,which achieves a better balance of speed and accuracy and has a larger perceptual field more suitable for object detection.Secondly,object detection requires identifying and localizing objects from the input image,including both classification and localization tasks.And the two focus on different features,while these features are usually located in different regions of the target.Therefore the classification and regression tasks are in some conflict.Most researchers currently utilize two independent branches to perform target classification and localization separately to solve the conflict problem.Since both branches are trained separately and cannot receive supervised information from each other,there is a lack of interaction and communication between them,resulting in inconsistency and misalignment between the two tasks at the time of prediction.In this paper,we propose a detection head(Attention Guide Task-aligned Head)with an attention mechanism to align the tasks.By introducing channel attention and spatial attention into the two-branch detection head,we can promote the alignment of the two tasks and improve the detection accuracy.Using the above lightweight convolutional neural network as the backbone network,combined with AGT-haed,this paper further designs a lightweight object detection called AT-YOLO.Experiments and evaluations on a large MS COCO(Microsoft Common Objects in Context)dataset confirm the model that AT-YOLO achieves 29.9% mAP with3.47 M parameters.in addition,our inference speed experiments on real devices show that ATYOLO achieves better results.Finally,an intelligent video surveillance system based on cloud-edge collaboration is designed to address the problems in the original central cloud-based architecture intelligent video surveillance system.The system uses Raspberry Pi 4B,an embedded computer,as the edge device side.The edge device performs video data acquisition,deploys the AT-YOLO object detection model for video processing,and finally uploads the detection results to the cloud.Considering the redundancy in the video stream and improving the detection efficiency,this paper also designs a key frame extraction algorithm using convolutional neural network(CNN)to extract features,which has only 0.32 M parameters and can accurately and quickly filter out key frames.Experimental results show that our intelligent surveillance system only needs to upload key frames to the cloud while meeting the real-time requirements,thus reducing bandwidth requirements.For 1080 P video,our system only needs 467.29 Kbps bandwidth to transmit key frames,while 5Mbps bandwidth is required to transmit the original video.