PaperlayerView exact revision on arXiv

Source-linked AI summary

Real-Time Human Detection as an Edge Service Enabled by a Lightweight CNN

Seyed Yahya Nikouei, Yu Chen, Sejun Song, Ronghua Xu, Baek-Young Choi, Timothy R. Faughnan

arXiv:1805.00330v1cs.CV

TL;DR

Large surveillance streams burden communication and storage, while video processing can exceed edge-device resources. The paper proposes a lightweight, human-focused CNN combining depthwise separable convolution with SSD and evaluates it on a Raspberry Pi 3. It reports an average speed of 1.79 FPS and a 6.6% false positive rate, presenting the approach as promising for smart surveillance.

  • Problem

    Surveillance video strains communication networks and storage, while video processing can be too computationally demanding for edge devices and human detection is a critical step in surveillance decision-making.

    Source: §I · p. 1, A. Smart Surveillance as an Edge Service · p. 2

  • Method

    The L-CNN combines depthwise separable convolution, human-focused network specialization, and SSD for object detection at the edge.

    Source: §III · p. 2, C. Human Object Detection Methods · p. 4

  • Results

    L-CNN averaged 1.79 FPS and had a 6.6% average false positive rate, compared with 26.3% for Haar Cascaded.

    Source: B. Results and Discussions · p. 4

  • Takeaways & Limitations

    The authors conclude that L-CNN is promising for smart surveillance, with reasonable processing speed and decent accuracy.

    Source: §V · p. 5

Abstract

from arXiv · show

Edge computing allows more computing tasks to take place on the decentralized nodes at the edge of networks. Today many delay sensitive, mission-critical applications can leverage these edge devices to reduce the time delay or even to enable real time, online decision making thanks to their onsite presence. Human objects detection, behavior recognition and prediction in smart surveillance fall into that category, where a transition of a huge volume of video streaming data can take valuable time and place heavy pressure on communication networks. It is widely recognized that video processing and object detection are computing intensive and too expensive to be handled by resource limited edge devices. Inspired by the depthwise separable convolution and Single Shot Multi-Box Detector (SSD), a lightweight Convolutional Neural Network (LCNN) is introduced in this paper. By narrowing down the classifier's searching space to focus on human objects in surveillance video frames, the proposed LCNN algorithm is able to detect pedestrians with an affordable computation workload to an edge device. A prototype has been implemented on an edge node (Raspberry PI 3) using openCV libraries, and satisfactory performance is achieved using real world surveillance video streams. The experimental study has validated the design of LCNN and shown it is a promising approach to computing intensive applications at the edge.

I. INTRODUCTION

Large, fast-growing surveillance streams strain communication networks and storage, while edge devices offer on-site computation for delay-sensitive surveillance. The paper introduces a lightweight human-detection CNN for edge devices and evaluates it on a Raspberry Pi 3 using real-world surveillance video.

I. INTRODUCTION

  • Surveillance video creates heavy communication and storage burdens; a 10 Hz WAMI sequence alone can exceed 250M data per second, and stored footage is often removed after weeks.The stated data rate can be greater for higher-frame-rate and higher-resolution video.Source: §I · p. 1
  • The paper proposes an L-CNN tailored to human detection and resource-constrained edge devices, with less preprocessing overhead than other human detection and classification algorithms.The introduction identifies real-time surveillance as a delay-sensitive, mission-critical use case for edge computing.Source: §I · p. 1
  • The L-CNN prototype averaged 1.79 FPS and reached 2.06 FPS on a Raspberry Pi 3, which the authors consider sufficient for detecting newly appearing pedestrians in real time.Previously detected objects are tracked separately, and objects that leave the frame are automatically deleted.Source: §I · p. 1
  • The authors implemented the L-CNN on a Raspberry Pi edge node, tested it with real-world surveillance video, and compared it with SSD GoogleNet, Haar-Cascade, and HOG + SVM.The introduction describes the experiments as validating the algorithm’s effectiveness and correctness for smart surveillance.Source: §I · p. 2

II. BACKGROUND KNOWLEDGE AND RELATED WORK

The background frames edge surveillance as a trade-off: local execution can reduce communication delays, but video processing remains demanding for edge devices. It motivates human-focused, lightweight CNNs and identifies MobileNet’s depthwise-separable design as the basis for the proposed network.

A. Smart Surveillance as an Edge Service

  • Moving functions to edge devices can minimize communication delays and message overhead, but video processing can burden edge devices enough to require decision-making at fog or cloud nodes.The text links missed human detections to false alarms and describes fog execution as near real time.Source: A. Smart Surveillance as an Edge Service · p. 2
  • Human detection is complicated by occlusion, illumination, and changing appearance, and accurate human-object coordinates are important for tracking.The discussion notes that detection algorithms may perform poorly when lighting, camera angle, or background texture differs from training and operation conditions.Source: B. Convolutional Neural Network · p. 2, A. Smart Surveillance as an Edge Service · p. 2

B. Convolutional Neural Network

  • The authors report that they could not find a deep-learning network designed specifically for human-object detection, and argue that human-only detection can use a network with fewer filters.General-purpose architectures such as GoogLeNet and ResNet classify many object types, including humans.Source: B. Convolutional Neural Network · p. 2
  • SqueezeNet reduces memory use through fire modules, while MobileNet separates regular convolution into depthwise and pointwise layers to reduce complexity and accelerate inference.MobileNet was selected as the base for the L-CNN; the passage compares its results with GoogleNet and ResNet,.Source: B. Convolutional Neural Network · p. 2

III. LIGHTWEIGHT CNN

The L-CNN is designed for edge devices by reducing convolutional computation and specializing the network for human detection. It combines this lightweight design with SSD for fast object detection and localization.

III. LIGHTWEIGHT CNN

  • Regular deep neural networks can require more parameter storage and RAM than edge devices can afford, even when their training is performed elsewhere.The passage gives ResNet architectures with 51, 101, or 201 layers as examples of networks with many parameters to load into small-device memory.Source: §III · p. 2
  • The L-CNN uses depthwise separable convolution to reduce CNN computation and specializes the network for human detection by reducing unnecessary filters.The authors state that this reduces computational cost without much sacrifice in overall accuracy.Source: §III · p. 2
  • The network combines a lightweight CNN with SSD to provide fast object detection and the pedestrian positions needed for movement-based decisions.The passage motivates accurate pedestrian localization as necessary for decisions based on movement.Source: §III · p. 2

A. Depthwise Separable Convolution

Depthwise separable convolution replaces a conventional convolution with depthwise and pointwise stages to reduce computational complexity. The paper presents this decomposition as a basis for making CNN computation more suitable for edge devices.

A. Depthwise Separable Convolution

  • Depthwise separable convolution splits regular convolution into depthwise and pointwise stages, making computational complexity more suitable for edge devices.The paper states that this approach’s computational-complexity reduction is calculated relative to conventional convolution [6].Source: A. Depthwise Separable Convolution · pp. 2–3
  • A conventional convolution maps an input with M channels to an output with N channels using Dk × Dk filters, with its computational complexity given by CCConventional.The input and output spatial dimensions are described as Df × Df and Dg × Dg, respectively.Source: A. Depthwise Separable Convolution · p. 3
  • The depthwise stage applies one Dk × Dk filter per input channel, while the pointwise stage uses N 1 × 1 filters.The figure compares conventional convolution with depthwise separable convolution.Source: A. Depthwise Separable Convolution · p. 3, Figure 1 · p. 3

B. The L-CNN Architecture

The L-CNN architecture has 26 layers before its classifier and output layers, using depthwise and pointwise convolutions through most of the network. It pools once and uses filter striding for downsizing.

B. The L-CNN Architecture

  • The L-CNN has 26 layers when depthwise and pointwise convolutions are counted separately, excluding the classifier, softmax, and regression layers.Source: B. The L-CNN Architecture · p. 3
  • A fully connected classifier uses prior probabilities for each proposed window to identify objects and label output bounding boxes.Source: B. The L-CNN Architecture · p. 3
  • Pooling occurs once at the end of the convolutional layers, while filter striding performs downsizing.Source: B. The L-CNN Architecture · p. 3

C. Human Object Detection Methods

The section describes proposal-based R-CNN methods and SSD’s unified detector. The proposed L-CNN uses SSD-style prior probability maps from several feature-map layers.

C. Human Object Detection Methods

A. Experimental Setup

The experimental setup uses a Raspberry PI 3 Model B as the edge device and records memory utilization while comparing algorithms.

A. Experimental Setup

  • Experiments use a Raspberry PI 3 Model B with an ARMv7 1.2 GHz processor and 1 GB of RAM.Source: A. Experimental Setup · p. 4
  • A memory profiler records utilization during algorithm execution, and SSD-GoogleNet and L-CNN use the same training databases.Source: A. Experimental Setup · p. 4
  • FPS performance is averaged over 30 seconds for Table I, and other CNNs must be converted to SSD for comparison.Source: A. Experimental Setup · p. 4

B. Results and Discussions

L-CNN combines competitive detection accuracy with higher speed than SSD GoogleNet, and the section compares its memory and performance with other CNNs and tests difficult viewing conditions. The reported accuracy measures use real-life surveillance video and may be higher than general-purpose results.

B. Results and Discussions

  • At 1.79 FPS average (2.06 peak), L-CNN records 6.6% FPR versus Haar Cascaded’s 26.3%, and accuracy comparable to SSD GoogleNet’s 5.3% FPR at 0.39 FPS.Table I covers FPS, CPU, memory utility, and average FPR.Source: Table I · p. 4, B. Results and Discussions · p. 4
  • Accuracy measures use real-life surveillance video, and the authors note they may be higher than results for general-purpose usage.The authors identify only one usage model for these calculations.Source: B. Results and Discussions · p. 4
  • L-CNN uses three times less memory than GoogleNet without a significant performance drop, and its fewer filters yield faster speed.The SSD-based GoogleNet used here was trained on 21 classes rather than the usual 1,000, reducing its memory use.Source: B. Results and Discussions · p. 4, Figure 3 · p. 4, Figure 4 · p. 4
  • Figure 5 tests varied angles and distances, including incomplete bodies; L-CNN handles such difficult cases that are often undetectable.The passage gives an example where both of a pedestrian’s legs are outside the frame.Source: Figure 5 · p. 5, B. Results and Discussions · p. 5
  • On a desktop, VGG requires up to 20 times more memory; L-CNN is faster than MobileNet and more accurate than SqueezeNet.The test allowed larger networks to be loaded; SqueezeNet is not SSD-compatible.Source: Table II · p. 5, B. Results and Discussions · p. 5

V. CONCLUSIONS

The paper introduces a lightweight CNN using depthwise separable convolution for human detection at the edge. It reports practical processing speed and false-positive performance, and describes handling people who are only partly visible in a frame.

V. CONCLUSIONS

  • At maximum 2.03 and average 1.79 FPS, L-CNN is comparable with Haar Cascaded and has a 6.6% false positive rate.The conclusion characterizes this as decent accuracy and reasonable processing speed for smart surveillance.Source: §V · p. 5
  • L-CNN uses depthwise separable convolution in a lightweight CNN for human-object detection at the edge.Source: §V · p. 5
  • L-CNN handles challenging situations where human objects are not completely in the frame.Source: §V · p. 5
Loading 1805.00330v1…