Table 3.

Video-based and hybrid localization techniques.

Ref (Year)	Technique	Algorithm Environment	Validation	Error/Accuracy	Notes
[77] (2022)	Image Processing with a self-attention mechanism	Attention Aware CNN	Public datasets: Bleeding dataset and Kvasir-Capsule dataset	Accuracy: Bleeding dataset: 95.1% Kvasir-Capsule dataset: 94.7%.	A dual-branch CNN model integrating self-attention mechanisms and using ResNet-50 to improve classification accuracy and lesion localization in WCE images at 30 fps.
[79] (2019)	Modified R-CNN	ResNet-50 and ResNet-101 models with data augmentation and fine-tuning	CVC-ColonDB, CVC-PolypHD, and ETIS-Larib	F1 score: 96.67% F2 score: 96.10%.	The work introduces a modified R-CNN for polyp identification and adapts deep learning models trained on non-medical images.
[80] (2021)	Deep CNN with attention mechanism	WCENet Grad-CAM++ and SegNet	KID dataset for WCE images	Accuracy: 98%, Dice Score: 56%	The study introduces a hybrid anomaly localization method for identification and segmentation of abnormal regions.
[83] (2021)	Feature point tracking techniques	SURF and RANSAC	84 videos from 42 patients	Error: 4 ± 0.7 cm	The study utilizes feature point tracking to estimate capsule displacement and orientation.
[84] (2021)	Hybrid: Video + IMU	Fusion Algorithm	Experiment: Ex-vitro porcine intestine	Accuracy: 0.95 cm	Hybrid method uses four low-resolution side-wall cameras and an IMU with a 9 DoF sensor for 6 DoF localization.
[85] (2022)	Hybrid: Video + RSSS + ToF	STN, HCO, CapsNet	Simulation: UWB, 8–50 RXs	Error: 5.41 mm Accuracy: 96.43%	The method integrates RF and vision-based data for localization using a fusion of multiple algorithms.
[86] (2022)	Hybrid: Video + Magnetic	MagnetO Fuse	Experiment: 3 × 3 sensor array, robotics arm, bio-tissues	Average Error: Stationary Capsule: 0.84 mm Moving Capsule: 3.5 mm	The proposed algorithm uses mathematical models to reconstruct the capsule’s position and low-resolution side wall cameras to assess motion.
[87] (2018)	Hybrid: Video + RSSI	CAC-RSSI, L-M	Experiment: Human mimicking phantom and pig small intestine	Error: 0.98 cm	A four-camera VGA-resolution WCE system is used to improve data transmission and localization accuracy, utilizing BCC, CAC-RSSI, and L-M algorithms.