1. Multi-Intelligence–Based Autonomous Driving

• Objective: Design an autonomous driving architecture that combines multiple AI modules instead of a single monolithic model.

• Main contents: Separate perception, localization, scene understanding, and driving decision modules for each sensor (camera, LiDAR, GNSS/INS).

• Characteristics: Compare the reliability of each module and select the most appropriate one depending on the situation.

• Goal: Realize explainable autonomous driving that can provide reasons for its decisions.

2. Intelligent LiDAR Point-Cloud SLAM and High-Precision Localization

• Objective: Achieve robust 3D SLAM and localization in complex environments.

• Main contents: Hybrid SLAM that combines geometry-based LIO-SAM and learning-based feature extractors (e.g., LCR-Net, DeLORA).

• Target environments: Repetitive or feature-poor urban, campus, and indoor spaces.

• Additional goal: Automatically generate high-resolution texture maps and HD maps using LiDAR intensity and signal information.

3. End-to-End Deep-Learning-Based Autonomous Driving (LiDAR/Camera/Radar Fusion)

• Objective: Integrate perception, prediction, and local planning into a single end-to-end network.

• Main contents: Design LiDAR–camera and radar–camera fusion architectures.

• Methods: Improve representation power and generalization via auxiliary tasks.

• Goal: Build and compare robust end-to-end driving models that perform reliably in bad weather and complex road conditions.

4. Automatic Visual Inspection (AVI) for Display Panels

• Objective: Develop an automatic defect detection system for OLED panels.

• Main contents: Design a deep-learning-based automatic visual inspection (AVI) network.

• Data: Handle low-label scenarios using multi-view information and semi-supervised learning.

• Goal: Achieve high detection accuracy and stable quality inspection for diverse defect types.

5. Passenger Posture and Behavior Recognition in Public Transport Vehicles

• Objective: Prevent safety accidents (falls, collisions, etc.) inside buses and other passenger vehicles.

• Main contents: Real-time recognition of passenger posture and behavior using in-vehicle camera images.

• Data: Build a passenger pose dataset from real in-vehicle videos.

• Methods: Skeleton-based action recognition to detect risky behaviors such as walking, standing, holding, falling, and sitting.

6. TQC-Based DRL Autonomous Navigation (Sensor-to-Policy Local Planner)

• Objective: Develop a reinforcement-learning-based continuous-control local planner using TQC.

• Inputs: Ouster 3D LiDAR converted to 360° 2D LaserScan, odometry, and IMU.

• Methods: Truncated Quantile Critics with top-quantile truncation and automatic entropy tuning.

• Additional components: Prioritized replay and N-step returns for stable learning.

• Goal: Build a Sim2Real navigation pipeline from Gazebo and ROS2 to real environments with actual LiDAR.

7. Edge Device for Passenger State Recognition to Prevent Bus Accidents

• Objective: Recognize passenger states on-device in real time for accident prevention in buses.

• Environments: Accurately estimate joint positions and poses even when occluded by seats and handrails.

• Methods: Unified detection-and-pose network (RTM-Det + RTMO) with a root heatmap.

• Data: Construct a bus-specific pose dataset combining real and synthetic data.

• Goal: Achieve real-time action recognition (~30–35 ms) and early warning for dangerous behaviors.

8. AI-Based Smart Vehicle Fault Prediction Framework

• Objective: Enable early fault diagnosis of electric vehicle drivetrains (motors and gearboxes).

• Main contents: Design a real-time semantic segmentation–based anomaly diagnosis model.

• Target performance: Lightweight architecture that can run at 100 FPS or more on embedded platforms.

• Characteristics: Diagnose not only the presence of faults but also their type and location.

9. Outdoor Safety Monitoring for Special and Large Vehicles

• Objective: Prevent outdoor safety accidents around special and large vehicles during boarding, alighting, and driving.

• Main contents: Build a monitoring system that detects nearby objects such as passengers, pedestrians, and obstacles.

• Sensors: Use fisheye-camera-based custom datasets and distorted images.

• Methods: Apply small-object-oriented detection algorithms with distortion correction and post-processing to reduce background false positives.

10. Deep-Learning-Based Multi-Camera Auto-Calibration and Surround View

• Objective: Automatically calibrate multiple fisheye cameras around a vehicle and generate surround-view images.

• Intrinsic parameters: Automatically estimated using deep learning.

• Extrinsic parameters: Precisely calibrated via marker-based alignment.

• Training: Improve calibration robustness using teacher–student semi-supervised learning.

• Goal: Provide a 360° bird’s-eye surround view with minimal distortion via BEV-based stitching and blending.

11. Adverse-Weather Image Restoration and Enhancement

• Objective: Restore and enhance images degraded by adverse weather (rain, fog, snow, etc.).

• Main components: Synthetic weather generator, conditional diffusion model, and SFHformer network.

• Methods: Hybrid representation combining frequency (Fourier) and spatial (Transformer) domains.

• Metrics: Improve PSNR, SSIM, and downstream perception performance in autonomous driving.

12. Small Object Detection and License Plate Recognition (ICR-Loss–Based)

• Objective: Improve detection performance for difficult small targets such as tiny objects and license plates.

• Core idea: Enhance learning via a loss function (ICR Loss) without changing the network architecture.

• Methods: Design a loss that models inter-class relations (Inter-Class Relation).

• Data: Validate performance with new datasets such as SVMLP.

• Extensibility: Expand to multi-sensor anomaly detection and perception for autonomous driving (camera–LiDAR–radar).

13. Automatic Post-Processing Criteria for OLED Auto-Labeling

• Objective: Improve the quality and fully automate post-processing of auto-labeling results on OLED production lines.

• Main contents: Develop the FASP (Fully Automated Selective Prediction) model.

• Architecture: Encoder–decoder that estimates label confidence.

• Function: Automatically determine post-processing thresholds that balance conservativeness and coverage without human intervention

• Effect: Reduce label noise and lower the cost of building manufacturing datasets.

14. Voxel-Based 3D Object Detection Framework

• Objective: Simultaneously improve accuracy and efficiency of LiDAR-based 3D object perception for autonomous driving.

• Main contents: Design a next-generation 3D detector that combines voxelization with spatial encoding.

• Target performance: Achieve or surpass state-of-the-art levels in both accuracy and computational efficiency.

• Application environments: Support real-time inference on large-scale autonomous driving datasets and in complex urban scenes

15. Speech Emotion Recognition for Autonomous Vehicles

• Objective: Enhance understanding of passenger state and driving context through speech-based emotion recognition.

• Main contents: Reconstruct the SER pipeline using state-of-the-art speech processing and deep-learning techniques.

• Goal: Reduce computational cost while improving emotion recognition accuracy.

• Applications: Serve as a core module for multimodal affective and situational understanding inside autonomous vehicles.