PaperlayerView exact revision on arXiv

Source-linked AI summary

DroNet: Efficient convolutional neural network detector for real-time UAV applications

Christos Kyrkou, George Plastiras, Stylianos Venieris, Theocharis Theocharides, Christos-Savvas Bouganis

arXiv:1807.06789v1cs.CV

TL;DR

UAV vehicle detection must operate under tight onboard power and latency constraints, while cloud offloading can add transmission delay. The paper develops and evaluates a lightweight single-shot CNN detector for embedded UAV processing. DroNet is reported to achieve 40× speed-up over previous works at ∼95% accuracy.

  • Problem

    UAV vehicle detection requires low-power, low-latency processing, while GPU batching and cloud video offloading can impose latency or power costs.

    Source: §I · p. 1

  • Method

    The paper adapts and explores single-shot CNN designs for top-view vehicle detection, aiming for efficient onboard execution on embedded UAV hardware.

    Source: A. Related Work on UAV-based Vehicle… · p. 2, C. CNN Models · p. 3

  • Results

    40× speed-up over previous works at ∼95% accuracy is reported for the proposed vehicle detector, DroNet.

    Source: §I · p. 2

  • Takeaways & Limitations

    DroNet is presented as a starting point for developing UAV-based object detectors for various applications.

    Source: §I · p. 2

Abstract

from arXiv · show

Unmanned Aerial Vehicles (drones) are emerging as a promising technology for both environmental and infrastructure monitoring, with broad use in a plethora of applications. Many such applications require the use of computer vision algorithms in order to analyse the information captured from an on-board camera. Such applications include detecting vehicles for emergency response and traffic monitoring. This paper therefore, explores the trade-offs involved in the development of a single-shot object detector based on deep convolutional neural networks (CNNs) that can enable UAVs to perform vehicle detection under a resource constrained environment such as in a UAV. The paper presents a holistic approach for designing such systems; the data collection and training stages, the CNN architecture, and the optimizations necessary to efficiently map such a CNN on a lightweight embedded processing platform suitable for deployment on UAVs. Through the analysis we propose a CNN architecture that is capable of detecting vehicles from aerial UAV images and can operate between 5-18 frames-per-second for a variety of platforms with an overall accuracy of ~95%. Overall, the proposed architecture is suitable for UAV applications, utilizing low-power embedded processors that can be deployed on commercial UAVs.

I. INTRODUCTION

UAVs can support traffic monitoring and emergency response, but onboard vehicle detection must meet tight power and latency constraints. The paper develops DroNet, a CNN architecture intended to detect vehicles efficiently on embedded UAV hardware.

I. INTRODUCTION

  • UAVs can provide vehicle information for traffic regulation and search-and-rescue, while supporting vehicle identification without sensors embedded in cars.The introduction also notes UAVs’ remote-sensing and situational-awareness benefits in emergency response.Source: §I · p. 1
  • UAV computing must minimize power use and latency, while cloud offloading can add video-transmission delay and may be unavailable without internet access.These constraints motivate efficient local processing on the UAV.Source: §I · p. 1
  • 40× speed-up over previous works with ∼95% accuracy is the stated target for the proposed vehicle detector, DroNet.The paper presents DroNet as a starting point for UAV-based object detectors.Source: §I · p. 2

II. BACKGROUND AND RELATED WORK

Prior UAV vehicle-detection approaches include traditional image-processing methods and CNN pipelines that can be inefficient or require high-end hardware. This work targets lightweight, onboard single-shot detection, against a background of CNN detector designs balancing accuracy and computational cost.

A. Related Work on UAV-based Vehicle Detection

  • Earlier UAV vehicle detection used background subtraction or Haar cascades, while CNN approaches could require thousands of search windows or offline high-end GPU processing.The paper identifies these constraints as unsuitable for lightweight, low-power UAVs.Source: A. Related Work on UAV-based Vehicle… · p. 2
  • The paper targets onboard processing with a lightweight CNN, applying single-shot detection trained specifically for top-view vehicles and optimized for embedded UAV hardware.The authors note that such detection techniques had not yet been exploited in UAV applications.Source: A. Related Work on UAV-based Vehicle… · p. 2

B. Convolutional Neural Network Detectors

  • CNN object detectors are categorized as region-based systems with proposal and classification stages or single-shot systems that perform end-to-end detection with one CNN.Detector design emphasizes accuracy and computational complexity for real-time mobile and embedded use.Source: B. Convolutional Neural Network… · p. 2

1) Region-based Detectors:

The background contrasts region-based detection pipelines with single-shot approaches designed to reduce processing bottlenecks. The paper focuses on YOLO because its portable C-based release supports evaluation across computing platforms.

1) Region-based Detectors:

  • Faster R-CNN proposes class-agnostic boxes, then classifies and refines them using cropped features; processing hundreds of proposals makes it computationally heavy for embedded platforms.The two-stage pipeline uses a region proposal network followed by a classifier.Source: 1) Region-based Detectors · p. 2

2) Single-Shot Detectors:

  • YOLO uses one CNN to predict bounding boxes, confidence, and class probabilities from image-grid cells, favoring real-time execution over detection accuracy relative to Faster R-CNN.The paper selects YOLO as its basis because its portable C-based release enables exploration across diverse platforms.Source: 2) Single-Shot Detectors · pp. 2–3
  • SSD extends YOLO with additional convolutional layers and wider box-grid aspect ratios to improve accuracy across multiple object scales.The paper focuses on single-shot detectors for their performance and applicability to mobile and embedded systems.Source: 2) Single-Shot Detectors · p. 2

III. PROPOSED APPROACH

The proposed approach collects varied UAV and other image sources, annotates visible vehicles for training, and explores CNN designs for efficient top-view vehicle detection. Its model exploration adapts Tiny-YOLO and varies network structure and input configuration.

A. Data Collection

  • 350 images containing approximately 5,000 vehicles were assembled from satellite-image crops, web images, and UAV traffic video to cover varied illumination, viewpoint, occlusion, color, and vehicle types.The dataset was intended to expose the detector to vehicles under different conditions.Source: A. Data Collection · p. 3

B. Training Process

  • Training annotations included vehicles with at least 50% of their body visible, and the Darknet-trained models jointly predicted vehicle presence and bounding boxes.Models were trained on an NVIDIA Titan Xp GPU and varied in layers, filter sizes, and input image size.Source: B. Training Process · p. 3

C. CNN Models

  • The CNN exploration adapted Tiny-YOLO to detect top-view vehicles and compared four designs by varying filters, layers, image size, convolution, and pooling structure.The four designs were SmallYoloV3, TinyYoloVoc, TinyYoloNet, and DroNet; the paper frames specialized UAV designs as addressing constrained compute and memory bandwidth.Source: C. CNN Models · p. 3

1) Number and Size of Filters:

The design varies network capacity and input resolution to balance detection speed, accuracy, and resource use; altitude-based size filtering is identified as complementary future work.

1) Number and Size of Filters:

  • The models start from Tiny-YOLO, reduce filters and prune layers to lower computation, and use nine convolutional layers with four to six max-pooling layers.Filter counts gradually increase with network depth to keep compute requirements low.Source: 1) Number and Size of Filters · p. 3

2) Input Image Size:

  • Input images were varied from 352 to 608 in training and testing because larger images can improve accuracy while increasing computational and memory load.Larger images include more information and create larger feature maps to process.Source: 2) Input Image Size · p. 3

D. Application Level Optimizations

  • Altitude-based filtering could discard detections whose sizes are infeasible for the UAV altitude and real vehicle size, but this optimization is left for future work.The approach applies when the UAV platform can provide altitude information.Source: D. Application Level Optimizations · p. 3

IV. EVALUATION AND EXPERIMENTAL RESULTS

The evaluation compares four CNN structures across input sizes and three computing platforms. It defines IoU and sensitivity as measures of detection quality.

IV. EVALUATION AND EXPERIMENTAL RESULTS

  • Four CNN structures are quantitatively evaluated on the vehicle dataset across input sizes and Intel i5-2520m, Odroid UX4, and Raspberry Pi 3 platforms.The selected platforms represent differing deployment configurations, from lightweight on-board systems to dedicated ground-station infrastructure.Source: §IV · p. 3

1) Intersection over Union (IoU):

  • For object detection, IoU captures the similarity between a predicted region and the ground-truth region for an object present in the image, defined as intersection size divided by their union.Source: 1) Intersection over Union (IoU) · p. 3

2) Sensitivity:

  • Sensitivity is the proportion of true positives correctly identified, calculated using true positives and false negatives.Source: 2) Sensitivity · p. 4

3) Precision:

The section defines precision and frames per second, then combines accuracy and speed measures into a weighted score. The score weights prioritize FPS for real-time applications.

3) Precision:

  • Precision is the proportion of true positives among all system detections.Source: 3) Precision · p. 4

4) Frames-Per-Second (FPS):

  • The composite score combines IoU, sensitivity, precision, and FPS, with FPS weighted 0.4 and each accuracy metric weighted 0.2.The authors describe the weights as reflecting each metric’s application-level importance and use the score to assess detector fit to platform constraints.Source: 4) Frames-Per-Second (FPS) · p. 4

A. Parameter Space Exploration on CPU Platform

The CPU-platform comparison examines how network architecture and input size affect detection performance. DroNet is substantially faster than TinyYoloVoC at 386-pixel inputs, with relatively small reported accuracy decreases.

1) Number and Size of Filters:

2) Input Image Size:

Larger input images can improve detection accuracy but reduce speed. For DroNet, the selected 512×512 input maximized weighted Score in the explored design space.

2) Input Image Size:

  • 97%, TinyYoloVoC with larger inputs achieved the highest test-set accuracy, while running 3× slower and gaining 0.17 IoU and 10% sensitivity.Source: 2) Input Image Size · p. 5
  • 1.28× average sensitivity gain and 0.81× average performance across models accompanied larger inputs; sizes 416–544 offered a favorable accuracy-speed trade-off.Source: 2) Input Image Size · p. 5
  • 512×512 maximized DroNet's weighted Score in the explored design space and was selected for UAV deployment.Source: 2) Input Image Size · p. 5

B. DroNet on UAV Platform

DroNet was deployed on a DJI Matrice 100 using an Odroid-XU4. The board achieved around 95% accuracy at 8–10 FPS, and the paper reports a 40× speedup over TinyYoloVoc on Odroid.

B. DroNet on UAV Platform

1) Odroid-XU4:

  • 40× faster than TinyYoloVoc, which achieved only 0.1 FPS on Odroid, DroNet ran at around 8–10 FPS while maintaining around 95% accuracy.Source: 1) Odroid-XU4 · p. 6

2) Raspberry Pi 3 - Model B:

On Raspberry Pi 3, DroNet maintained around 95% accuracy at 5–6 FPS, which the authors consider viable for applications with slow-varying dynamics. The conclusion reports 5–18 FPS across platforms and identifies precision and training-set expansion as future work.

2) Raspberry Pi 3 - Model B:

V. CONCLUSION & FUTURE WORK

  • 5–18 FPS across platforms at 95% detection accuracy summarizes the reported DroNet performance.Source: §V · p. 6
  • Future work includes finer-level bitwidth optimizations and expanding the training set with additional images and object classes.Examples include pedestrians and motorbikes for applications such as emergency response, traffic management, and infrastructure inspection.Source: §V · p. 6
Loading 1807.06789v1…