Object Detection in Adverse Weather Conditions using Tightly-coupled Data-driven Multi-modal Sensor Fusion

Supervisors:
- Prof. Dr. Ing- Sebastian Houben
- M.Sc. Santosh Thoduka

Motivation

Why multi-modal sensor fusion?


Figure 1: Sensors modality characteristics	Figure 2: Sensors modality characteristics

Report and Presentation:

R&D detailed report link (required access): KevinPatelRnD_Report
R&D Presentation (pdf): KevinPatelRnD_Presentation

Datasets

Table 3.1: Multimodal adverse weather conditions datasets. Sensors†: C-R-L-N-F denotes Camera, Radar, LiDAR, Near-infrared, and Far-infrared sensors, respectively. Weather conditions‡: F-SN-R-O-SL-N denotes Fog, Snow, Rain, Overcast, Sleet, and Night conditions, respectively. Note that DENSE and nuScenes datasets are used for the project.

Sorted in ascending order w.r.t year column

Name	Sensors†	Weather Cond.‡	Size (GB)	Year	Citation Cnt.	Link	Publisher	Pros	Cons
DENSE	CRLNF	F, SN, R, N	582	2020	269	Link	Mercedes, Ulm, Princeton	- More adverse weather data - Higher resolution data than nuScenes	- Less label frames - Sparse Radar data
nuScenes	CRL	R, N	400	2020	3459	Link	Motional	- Well documented - Heavily used	- Not good for adverse weather conditions - Sparse Radar data
The Oxford RobotCar	CRL	R, SN, F	4700	2020	317	Link	Oxford Robotics Institute
EU Long-term	CRL	SN, R, O, N		2020	72	Link	University of Technology of Belfort-Montbéliard (UTBM)
RADIATE	CRL	F, SN, R, O, SL, N		2021	132	Link	Heriot-Watt University
K-Radar	CRL	F, R, SN	13000	2022	15	Link	KAIST University	- Includes 4D radar	- Heavy dataset, order via physical drive
Boreas	CRL	SN, R, O, N	4400	2022	38	Link	University of Toronto	- High resolution radar
aiMotive	CRL	R, O, N	85	2023	3	Link	aiMotive company	- Fog and Snow not included (future work) - Relatively small dataset

Methods

Table 3.4: Multi-modal sensor fusion methods. Sensors†: C-R-L denote Camera, Radar, and LiDAR sensors, respectively

SAF-FCOS, HRFuser, and MT-DETR methods are thoroughly analyzed in the report

Name	Sensors†	Dataset Used	Fusion method	2D/3D	Code Link	Year	Published at	Cited By	Comment 1	Comment 2	Framework
CRF Net	CR	nuScenes	Data-level	2D	Link	2019	SDF	208	Uses BlackIn method for training	didn't mention NDS	Tensorflow
SAF-FCOS	CR	nuScenes	Feature-level	2D	Link	2020	Sensors	105	New spatial fusion strategy	AP = 72.4, didn't mention NDS	PyTorch
BIRANet	CR	nuScenes	Feature-level	2D	Link	2020	ICIP	36			PyTorch
GRIF Net	CR	nuScenes		NA	NA	2020					NA
SeeingThroughFog	CRLNF	DENSE	Feature-level	2D	NA	2020	CVPR	236	Novel entropy based net	normal to adverse weather transfer	NA
YOdar	CR	nuScenes		2D	NA	2020	ICAART	23			NA
CenterFusion	CR	nuScenes	Feature-level	3D	Link	2021	WACV	170	data augmentation applied	NDS = 44.0	PyTorch
RODNet	CR	CRUW	Feature-level	2D	Link	2021	WACV	58	Uses unique Radar data processing		PyTorch
CRAMNet	CR	RADIATE		NA		2022					NA
Attention Powered- #1	CR	nuScenes		NA	NA	2022					NA
Attention Powered- #2	CR	RADIATE		2D	NA	2022	CISDS	0	Outperform SAF-FCOS, CenterFusion		NA
MT-DETR	CRL	DENSE	Mixed-level	2D	Link	2023	WACV	2	Attention based method		PyTorch
RTNH	R	K-Radar		3D	Link	2023	NeurIPS	9	Baseline method uses only radar	4D radar dataset with AW	PyTorch
HVDetFusion	CR	nuScenes		3D	Link	2023		2	NDS = 67.4, built on top of CenterFusion		PyTorch
REDFormer	CR	nuScenes		3D	Link	2023	ITSC	0	NDS = 48.6, multi camera input, BEV based	how did they define SOTA in low visibility subset?	PyTorch
RADIANT	CR	nuScenes	Feature-level	3D	Link	2023	AAAI	2	didn't mention NDS	How come this is SOTA?	PyTorch
HRFuser	CRL	nuScenes, DENSE	Mixed-level	2D	Link	2023	ITSC	8	Mixed Fusion, Transformer based	didn't mention NDS	PyTorch
CamRaDepth	CR	nuScenes			Link	2023		Not yet published			PyTorch
AutoFed	CRL	The Oxford RobotCar		NA		2023					NA
aiMotive	LR	aiMotive		3D	Link	2023	ICLR	2	Yet to explore for Camera+Radar fusion		PyTorch

Figures from the report:

A few sample figures highlighting the importance of multi-modal sensor fusion


Figure 3: Van occluded by a water droplet on the lens


Figure 4: LiDAR performance test


Figure 5: 1st row: clear weather condition, 2nd row: with fog. Shows that lidar affects by the fog but radar intensity remains the same


Figure 6: Highlighting the significance of fusing multimodal sensor data


Figure 7: Samples of K-Radar datasets for various weather conditions

TODOs:

Contact:

Email 📧: kevinpatel4400@gmail.com

Citation:

@unpublished{RnDPatel,
    abstract = {In the ﬁeld of autonomous vehicles, object detection is a critical component, especially in perceiving the environment under adverse weather conditions. Traditional methods, primar- ily focused on camera data, face signiﬁcant limitations in such scenarios. This research aims to address these challenges through the exploration of multimodal sensor fusion, incor- porating Cameras, LiDAR, and Radar, to improve detection accuracy in inclement weather. The study primarily focuses on a tightly-coupled fusion approach, contrasted against the existing middle fusion strategy, with experiments conducted using the nuScenes and DENSE datasets, the latter featuring extreme weather conditions. The ﬁndings indicate that the integration of complementary sensors substantially enhances detection accuracy across various weather conditions and that the tightly-coupled fusion approach outperforms the middle fusion method. Both qualitative and quantitative analyses support these conclusions, highlighting the eﬀectiveness of this approach in the advancement of object detection technologies in autonomous vehicles. This research provides signiﬁcant insights into the robustness of sensor fusion techniques, oﬀering substantial contributions to the ﬁelds of computer vision and autonomous vehicle technology.},
    title = {Object detection in adverse weather conditions using tightly-coupled data-driven multi-modal sensor fusion},
    author = {Patel, Kevin},
    year = {2023},
    month = {December},
}

kvnptl/KevinPatelRnD