Lightweight Low-Altitude UAV Object Detection Based on Improved YOLOv5s

Addressing redundant computation, Chen et al. [14] introduced Partial Convolution (PConv), which reduces memory access while optimizing the parameter problem caused by redundant computation, greatly improving the ability to capture spatial features. FasterNet is constructed with PConv and 1×1 convolutional structures. In Figure 2, h, w, and k represent the height, width, and kernel size of the feature map, respectively, while Cp indicates the number of channels in conventional convolution.

Figure 2.

PConv only uses general Conv to achieve spatial feature acquisition on some input channels, while maintaining the remaining channels unchanged. Calculate by considering the first or last consecutive Cp channel as a representation of the entire feature map. Ensure its generality while maintaining the same number of input and output feature map channels. The FLOPs of PConv are h × w × k2 × C2p, which only accounts for 1/16 of the general Conv.

The C3 module in the YOLOv5 network is pivotal for increasing network depth and receptive field, enhancing feature extraction. Initially, it included Conv1, Conv2, Conv3 modules, and one or more Bottleneck modules. While this design enriches the learning capabilities of the C3 module, it also adds to the computational load and model complexity. Thus, this paper introduces a lightweight feature extraction module, C3F, inspired by the FasterNet module concept.

Figure 3 illustrates the structural diagram of the C3F feature extraction module. Here, h, w, and k signify the height, width, and convolution kernel size of the feature map, while Cp represents the number of conventional convolution channels.This module introduces partial convolution and replaces BattleNet with FasterNet Block in the C3 module, reducing computational redundancy and memory access while maintaining the speed and efficiency of feature extraction.

Figure 3.

Object detection models often employ nearest neighbor or bilinear interpolation for feature map upsampling. While adaptive upsampling uses methods such as deconvolution. However, these traditional methods have certain shortcomings in accurately reconstructing target detail information, which can easily lead to partial information loss of small targets, thereby affecting detection accuracy. In contrast, CARAFE improves the quality of upsampled features by recombining content aware features to make the upsampling kernel semantically relevant to the feature map. The CARAFE operator can better preserve and recover feature information details, so this article chooses to use the CARAFE operator for upsampling to enhance regional sensitivity and generate more accurate high-resolution feature maps.

Figure 4.

In various computer vision tasks, the significant effectiveness of channel or spatial attention mechanisms in generating clearer feature representations has been demonstrated. However, modeling cross channel relationships through channel dimensionality reduction may have side effects on extracting deep visual representations.

EMA without dimensionality reduction [16] aims to preserve channel information while minimizing computational overhead. This is achieved by reshaping some channels into batch dimensions and grouping channel dimensions into multiple sub-features to evenly distribute spatial semantic features. EMA learns effective channel descriptions in convolutional operations without reducing channel dimensionality, enhancing pixel-level attention for advanced feature maps. This lightweight and flexible EMA attention model serves as a core module applicable to lightweight networks.

Figure 5 outlines the EMA attention module, comprising Feature Grouping, Parallel Subnetworks, and Cross spatial learning. EMA divides the input feature map into multiple sub-features based on channel dimensions. It employs two parallel subnetworks: one with 1×1 branches and the other with 3×3 branches. The 1×1 branch encodes channel information using global average pooling operations, while the 3×3 branch captures local cross-channel interaction. EMA then fuses the output feature maps of the two subnetworks using cross-space learning, resulting in an attention weight map of the same size as the input feature map to enhance its expressive power.

Figure 5.

Large deep learning models are difficult to deploy on industrial embedded devices. Many lightweight networks use a large number of depthwise separable convolutions, and even if the C3 of the backbone network is replaced with lighter modules, there are still a large number of 1x1 convolution operations, making it difficult to achieve sufficient accuracy. This dense convolution operation actually consumes more resources, and even with channel shuffling, the effect is still poor. Therefore, this article embeds the Slim Neck [17] network is integrated into the feature fusion stage, incorporating the GSConv module and the VOVGSCSP module.

The GSConv module consists of the Conv, DWConv, Concat, and Channel Mixing sub-modules, illustrated in Figure 6. Its operational procedure is as follows: Initially, the input feature map, with C1 channels, undergoes standard convolution to produce a feature map with C2/2 channels. Subsequently, depthwise separable convolution generates another feature map with C2/2 channels. These two feature maps are concatenated to form a unified feature map with C2 channels. Finally, the channel mixing operation adjusts the output characteristics to the desired channel count. Through this approach, the GSConv module combines depthwise separable convolution and standard convolution to reduce computational complexity and improve overall recognition accuracy by addressing limitations in feature extraction and fusion.

Figure 6.

The integration of GSConv convolution aims to simplify model complexity. To enhance model inference speed without sacrificing accuracy, we introduced the VOVGSCSP module, as depicted in Figure 7. In Figure 7, Unit 1 illustrates the bottleneck unit structure of VOVGSCSP, while Unit 2 showcases a cross-stage VOVGSCSP module employing a single aggregation method.

Figure 7.

Figure 8.

In the identical operational setting and employing the UAV dataset, experimental tests are conducted to compare the enhanced ATD-YOLO algorithm model with other algorithm models. The experimental environment configuration is outlined in Table 1. This comparison seeks to validate the improved algorithm’s efficacy in detecting low-altitude UAV targets.

Throughout the training phase, a combination of the cosine annealing learning rate decay method and the SGD algorithm is utilized. The training regimen spans 600 epochs with a batch size of 16 and a momentum of 0.937. Mosaic data augmentation is employed to enrich the backgrounds of images by randomly scaling, cropping, and arranging four training images in a mosaic pattern. This augmentation technique enhances the accuracy and robustness of small object detection.

In this study, extensive data from diverse sources and papers was reviewed and collected. Utilizing this data, we constructed a tailored dataset called “Anti-Mini Drone” for our research purposes. The Det-Fly dataset [18] addresses the lack of drone data from a single perspective by directly collecting images of target drones in the air, including various postures such as upward, downward, and forward views. However, this dataset only contains one type of drone, limiting its generality in detecting other types of drones. The Drone-vs-Bird dataset [19] not only covers rich drone and environmental data but also includes some bird data. This poses challenges when drones resemble birds in appearance, especially during long-distance observations. However, the drawback of this dataset is its inability to meet the detection requirements of other types of drones. The Real World dataset [20] contains various types of drones and environments sourced from YouTube videos, but the image resolution is low. Most of the data is captured from a forward and upward perspective, which implies certain limitations in drone detection from a downward view. The Multi-view drone tracking dataset [21] records drone flight trajectories from different angles using multiple consumer-grade cameras, but the environmental capture is relatively homogeneous. The DUT anti-UAV dataset [22] consists of both a detection dataset comprising 10,000 images and a tracking dataset containing 20 videos. However, it’s worth noting that the distribution of target dimensions within the dataset is uneven. The Anti-UAV dataset [23] includes visible light and infrared data, but the problem is that the shooting environment is singular, suitable only for research on multi-modal fusion tracking and the alignment between infrared and visible light cameras is not perfect in time and space.

Overall, the above drone datasets each have their own advantages and disadvantages, often overcoming only one or two difficulties in constructing drone datasets. By merging these six different datasets, a low-altitude drone dataset that meets the requirements of this study was constructed, making it more reflective of real outdoor flight scenarios for low-altitude drones. Table 2 illustrates the distribution of images.

In the Anti-Mini Drone dataset, Figure 9 demonstrates a notable prevalence of small targets. The majority of targets in the dataset exhibit aspect ratios less than 0.1 times the original image dimensions. This distribution aligns with the relative sizes of objects commonly encountered in real outdoor scenarios during low-altitude unmanned aerial vehicle (UAV) flights.

Figure 9.

To evaluate the enhanced ATD-YOLO algorithm, metrics such as parameter count, floating-point operations (GFLOPs), average precision (AP), and frames per second (FPS) are selected. Since the study focuses on detecting drones across different categories, mean average precision (mAP) and AP values are considered equivalent. Given the predominance of small objects, the mAP.5 criterion is adopted for evaluation to reflect the model’s performance and speed accurately. Precision measures the proportion of correctly detected objects, while recall assesses the proportion of correctly predicted objects among all true objects: 1 Precision=TPTP+FP 2 Recall=TPTP+FN

FN represents false positive samples predicted by the model, TP denotes true positive samples predicted correctly, and FP stands for false negative samples. The average precision (AP) reflects the detection accuracy for a single class of targets, usually calculated by integrating the Precision-Recall (P-R) graph: 3 AP=∫01P(R)dR

Detection speed is frequently measured in FPS (Frames Per Second), indicating the number of images processed by the object detection network per second. A higher FPS value signifies faster processing speed. The expression for FPS is given by Formula 4: 4 FPS=FrameNumElapsedTime

Figure 10.

Figure 11.