RA-CRPN: Method for Detecting Small Distant Objects in Road Vehicle Vision

Fig. 7. Residual attention module and its constituent modules. (a) Residual attention module; (b) residual block; (c) mask branch; (d) feature fusion module

Download full size

Fig. 8. Improved CRPN model and constituent modules. (a) Improved CRPN model diagram; (b) GAM attention mechanism diagram

Download full size

Fig. 9. Visualization of detection for each target category by each model. (a) SODA-D schematic diagram; (b) realistic scene diagram; (c) target category diagram

Download full size

Table 1. Comparison of detection indicators for various models

View table

Table 1. Comparison of detection indicators for various models

Method	Schedule	FPS /frame/s	AP /%	AP@50 /%	AP@75 /%	AP_es /%	AP_rs /%	AP_gs /%	AP_n /%
One-stage
FCOS^［15］	12	32.8	23.8	49.7	20.0	7.1	19.6	31.1	41.2
YOLOX-M^［7］	50	50.9	26.9	54.1	23.0	13.8	25.0	31.2	30.6
YOLOv10-M^［8］	50	84.6	31.2	61.4	27.1	15.1	28.0	36.4	44.5
Keypoint-based
CornerNet^［30］	24	16.3	24.7	49.5	21.8	6.3	20.7	32.5	43.9
RepPoints^［31］	12	18.7	27.8	55.8	24.8	10.3	23.7	35.3	45.3
Query-based
Deformable-DETR^［32］	50	27.3	19.3	44.7	13.9	6.4	15.3	25.0	34.3
Sparse R-CNN^［33］	12	27.6	24.4	50.4	20.5	9.0	20.5	30.2	39.6
Two-stage
Baseline^［9］	12	8.6	28.8	59.5	24.1	13.9	25.8	34.3	43.1
Cascade RPN^［10］	12	8.2	29.2	56.6	26.1	12.6	25.7	35.6	44.3
RFLA^［34］	12	9.9	29.9	60.4	25.1	13.4	26.8	35.6	44.2
CFINet^［11］	12	8.7	30.8	60.9	26.9	14.8	27.6	36.3	44.6
Proposed	12	9.5	32.7	61.9	27.1	16.3	29.2	37.4	44.8

Table 2. Comparison of proposed generation methods and results

View table

Table 2. Comparison of proposed generation methods and results

Proposal method	AR /%	AR_es /%	AR_rs /%	AR_gs /%	AR_n /%
RPN	41.3	24.1	38.3	47.2	57.2
RPN-0.5	41.3	24.2	38.5	47.3	54.3
GA-RPN^［4］	42.2	24.1	39.3	48.8	56.3
Cascade RPN	41.9	22.8	28.3	48.7	57.1
CRPN	42.7	24.6	38.8	49.0	56.8
Improved CRPN	42.9	25.1	39.0	49.0	56.9

Table 3. Results of ablation experiment table

View table

Table 3. Results of ablation experiment table

Network				Parameters /10⁶	FLOPs /Gbit	AP /%	AP_es/%	AP_rs /%	AP_gs /%
Baseline	RO-ResNet	RA-ResNet	Improved CRPN	Parameters /10⁶	FLOPs /Gbit	AP /%	AP_es/%	AP_rs /%	AP_gs /%
√				44.5	198.4	29.0	13.6	25.8	34.1
√	√			50.2	213.3	30.7	14.8	27.6	36.4
√		√		45.5	200.6	30.8	14.8	27.6	36.5
√			√	54.3	243.1	31.3	14.9	27.8	36.4
√	√	√		51.2	215.5	32.3	15.7	28.7	37.1
√	√		√	60.0	258.0	31.5	15.1	28.1	36.6
√		√	√	55.3	245.3	31.4	15.3	28.1	36.7
√	√	√	√	61.0	260.2	32.7	16.3	29.2	37.4

Table 4. Ablation experiments of ODC under different modules and algorithms unit: percentage points

View table

Table 4. Ablation experiments of ODC under different modules and algorithms unit: percentage points

Module	Difference without ODC				Method	Difference without ODC
Module	AP	AP_es	AP_rs	AP_gs	Method	AP	AP_es	AP_rs	AP_gs
RO-ResNet	0.7	0.8	1.1	1.2	YOLOv10-M	1.5	1.2	1.3	1.3
RA-ResNet	0.7	0.9	1.2	1.1	RepPoints	1.2	1.2	1.4	1.5
ODC-FPN	0.9	1.0	1.2	1.2	Sparse R-CNN	1.3	1.3	1.5	1.4
‒	‒	‒	‒	‒	CFINet	2.3	1.8	1.9	2.1

Table 5. Influence of hyperparameters $α_{3}$ on the simulation of FI branch characteristics
View table
Table 5. Influence of hyperparameters $α_{3}$ on the simulation of FI branch characteristics
$α_{3}$ AP /% AP_es /% AP_rs /% AP_gs /%
0.25 31.3 15.2 27.9 36.5
0.50 31.7 15.3 28.2 36.8
0.75 31.1 14.7 28.0 36.5

Table 6. Detection results of each model for each target category in custom dataset

View table

Table 6. Detection results of each model for each target category in custom dataset

Parameter	Model	Faster R-CNN	Cascade RPN	RFLA	CFInet	RA-CRPN
AP@50 /%	People	70.2	64.3	73.5	74.8	75.2
	Rider	43.2	40.4	47.3	47.6	48.5
	Bicycle	54.6	50.3	60.1	61.9	63.4
	Motor	50.3	49.6	55.1	55.2	56.3
	Vehicle	63.4	60.1	67.6	68.2	69.1
	Traffic-light	53.8	50.2	56.2	56.9	58.1
	Traffic-sign	60.5	59.1	64.2	64.3	66.5
mAP@50 /%		56.6	53.4	60.6	61.6	62.4

Tools

Get Citation

Copy Citation Text

Xiaowei Xu, Jianyu Li, Qinghua Qi, Mingxing Deng. RA-CRPN: Method for Detecting Small Distant Objects in Road Vehicle Vision[J]. Laser & Optoelectronics Progress, 2025, 62(12): 1237005

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites

Paper Information

Category: Digital Image Processing

Received: Nov. 15, 2024

Accepted: Dec. 12, 2024

Published Online: Jun. 12, 2025

The Author Email: Mingxing Deng (dengmingxing@wust.edu.cn)

DOI:10.3788/LOP242269

CSTR:32186.14.LOP242269

Topics

laser devices and laser physics

Lasers and Laser Optics

Laser physics

laser manufacturing

Instrumentation, Measurement and Metrology

Table 1. Comparison of detection indicators for various models

Table 1. Comparison of detection indicators for various models

Table 2. Comparison of proposed generation methods and results

Table 2. Comparison of proposed generation methods and results

Table 3. Results of ablation experiment table

Table 3. Results of ablation experiment table

Table 4. Ablation experiments of ODC under different modules and algorithms unit: percentage points

Table 4. Ablation experiments of ODC under different modules and algorithms unit: percentage points

Table 5. Influence of hyperparameters α3 on the simulation of FI branch characteristics

Table 5. Influence of hyperparameters α3 on the simulation of FI branch characteristics

Table 6. Detection results of each model for each target category in custom dataset

Table 6. Detection results of each model for each target category in custom dataset

Table 5. Influence of hyperparameters $α_{3}$ on the simulation of FI branch characteristics

Table 5. Influence of hyperparameters $α_{3}$ on the simulation of FI branch characteristics