Lightweight target detection network for UAV platforms

Dandan HUANG; Han GAO; Zhi LIU; Lintao YU; Huiji WANG

doi:10.37188/OPE.20233120.3021

Optics and Precision Engineering, Volume. 31, Issue 20, 3021(2023)

Lightweight target detection network for UAV platforms

Dandan HUANG¹, Han GAO¹, Zhi LIU^1,2、*, Lintao YU¹, and Huiji WANG¹

¹School of Electronics and In formation Engineering， Changchun University of Science and Technology， Changchun30022， China

²National and Local Joint Engineering Ｒesearch Center of Space Photoelectric Technology， Changchun University of Science and Technology， Changchun1300， China

show less

Abstract Get PDF(in Chinese)

Figures & Tables(17)

Fig. 1. YOLOV5's Backbone layer various structural diagrams

Download full size

View in Article

Fig. 2. FasterNet network structure

Download full size

View in Article

Fig. 3. PConv working principle

Download full size

View in Article

Fig. 4. Comparison of convolutional variants

Download full size

View in Article

Fig. 5. Improved YOLOV5 model

Download full size

View in Article

Fig. 6. C3_FN network structure

Download full size

View in Article

Fig. 7. VisDrone 2019 Data set target statistical diagram

Download full size

View in Article

Fig. 8. Effect of ultra -parameter on YOLOV5 model

Download full size

View in Article

Fig. 9. Object detection schematic diagram of this model

Download full size

View in Article

Fig. 10. Exeriment diagram on Jetson Nano

Download full size

View in Article

Table 1. Setting value of the anchor frame of each detection branch

View table

View in Article

Table 1. Setting value of the anchor frame of each detection branch

检测分支	锚框设定
P2	（1，4），（2，9），（5，6）
P3	（5，13），（10，10），（8，20）
P4	（19，17），（15，31），（34，42）
P5	（30，61），（62，45），（59，119）

Table 2. Different detection branches comparison results
View table
View in Article
Table 2. Different detection branches comparison results
包含的检测分支 mAP_0.5/%
P3，P4，P5 40.8
P2，P4，P5 41.2
P2，P3，P5 40.7
P2，P3，P4 40.8
P2，P3，P4，P5 44.7

Table 3. Different predictive boundary box comparison results
View table
View in Article
Table 3. Different predictive boundary box comparison results
总训练轮数预测边界框 mAP_0.5/%
50 epochs CIOU 39.6
NWD 38.8
100 epochs CIOU 40.8
NWD 41.1

Table 4. Different network structure performance comparison
View table
View in Article
Table 4. Different network structure performance comparison
方法 mAP_0.5/% Parameters/M FLOPs/G Inference time/ms
C3 40.8 44.0 180.6 70.8
C3_Ghost 33.4 19.6 63.3 65.8
C3_FN 40.9 29.9 70.8 36.2

Table 5. VisDrone test data set experiment results

View table

View in Article

Table 5. VisDrone test data set experiment results

方法	mAP_0.5/ %	mAP_0.5-0.95/ %	Inference time/ms
Light-RCNN^［29］	39.5	23.2	52.1
Cascade-RCNN^［30］	37.8	22.56	57.3
RetinaNet^［31］	31.6	20.1	48.6
YOLOv5l	40.8	23.9	76.8
TPH-YOLOv5	46.2	27.5	55.1
YOLOv7	42.1	23.4	48.9
OURS	47.6	28.7	45.9

Table 6. VisDrone test data set discipline experiment
View table
View in Article
Table 6. VisDrone test data set discipline experiment
方法 mAP_0.5/% Parameters/M FLOPs/G Inference time/ms
Baseline 40.8 44.0 108.4 76.8
+P2分支 44.7 48.1 186.6 88.8
+NWD与IOU混合 47.2 48.1 186.6 89.7
+C3_FN 47.6 32.8 121.6 45.9

Table 7. Comparison of inference time before and after TensorRT acceleration
View table
View in Article
Table 7. Comparison of inference time before and after TensorRT acceleration
模型加速前平均推理时间加速后平均推理时间
YOLOv5l 563 153
OURS 342 84

Tools

Get Citation

Copy Citation Text

Dandan HUANG, Han GAO, Zhi LIU, Lintao YU, Huiji WANG. Lightweight target detection network for UAV platforms[J]. Optics and Precision Engineering, 2023, 31(20): 3021

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites