Autonomous Vehicles & AI Driving

Real-time detection and classification of vehicles, pedestrians, cyclists, and static obstacles using camera and LiDAR models (YOLO, PointPillars, CenterPoint) tuned for automotive latency budgets.

Kalman/particle-filter and deep-association tracking that maintains stable object IDs, velocity, and trajectory history across frames and sensor dropouts.

Early, late, and deep fusion architectures that combine LiDAR point clouds, radar returns, and camera imagery into a unified, calibrated perception output with graceful degradation.

Pixel- and point-level scene segmentation for drivable surface, road furniture, and free-space estimation using networks such as DeepLab, SegFormer, and Cylinder3D.

Robust lane-line, curb, and road-edge detection under occlusion, weather, and worn-marking conditions to feed lateral control and mapping.

Detection and state classification of traffic signs and signal lights, with regional sign taxonomy support and temporal filtering for reliable state estimation.

Vulnerable-road-user detection with pose, intent, and crossing-prediction models to support AEB and interaction-aware planning.

Intrinsic/extrinsic calibration and online calibration monitoring across camera, LiDAR, radar, and IMU to keep the fused frame geometrically consistent.

Simultaneous localization and mapping using LiDAR (LOAM/LIO-SAM) and visual-inertial (VINS, ORB-SLAM) pipelines for GPS-denied and unmapped environments.

Tightly coupled GNSS-RTK and inertial fusion with error-state Kalman filtering for high-rate, drift-corrected global positioning.

Dead-reckoning and visual-odometry estimation that bridges GNSS outages and provides continuous ego-motion for the localization stack.

Production of lane-level HD maps in OpenDRIVE/Lanelet2 formats, including semantic annotation, versioning, and change-detection workflows.

Point-cloud and feature registration against HD maps (NDT/ICP matching) for robust, centimeter-level pose within pre-mapped operational domains.

Fleet-sourced change detection and incremental map update pipelines that keep HD maps current with road works and lane reconfigurations.

Rule-based and learned behavior planners for lane changes, merges, intersections, and yielding, encoding traffic rules and right-of-way logic.

Sampling- and optimization-based trajectory generation (lattice, RRT*, Frenet) producing kinematically feasible, collision-free paths with comfort constraints.

Longitudinal and lateral MPC controllers that track planned trajectories while respecting actuator, tire, and stability limits.

Multi-agent trajectory prediction that anticipates the motion of surrounding vehicles and VRUs to enable interaction-aware planning.

Pure-pursuit, Stanley, and PID/LQR controllers integrated with drive-by-wire steering, throttle, and brake actuation.

Learned driving policies via imitation and reinforcement learning, benchmarked in simulation before shadow-mode and closed-course validation.

Radar/camera-fused longitudinal control maintaining set speed and time-gap headway, including stop-and-go and cut-in handling.

Lateral assist that keeps the vehicle centered in-lane with torque overlay steering and driver-hands-on monitoring.

Forward-collision detection and autonomous braking for vehicles, pedestrians, and cyclists, calibrated against Euro NCAP/IIHS test scenarios.

Corner-radar and surround-vision monitoring for blind-spot detection, lane-change assist, and rear cross-traffic alert.

Ultrasonic and surround-view parking assist, including perpendicular/parallel auto-park, valet parking, and low-speed summon maneuvers.

In-cabin camera systems for drowsiness, distraction, and gaze tracking plus occupant classification, aligned to GSR and NCAP DMS requirements.

L2+ hands-on/hands-off assist combining ACC, lane centering, and automated lane change for structured highway and congestion driving.

Architecture and integration of ROS 2 / DDS-based autonomy stacks with real-time QoS tuning, node lifecycle management, and deterministic messaging.

Deployment and customization of the open-source Autoware autonomous-driving stack across perception, planning, and control modules.

Integration and extension of the Apollo autonomous-driving platform, including Cyber RT runtime and dreamview tooling for target vehicles.

Porting and optimizing perception and planning workloads onto NVIDIA DRIVE Orin and Jetson platforms with the full CUDA/cuDNN pipeline.

Quantization, pruning, and TensorRT engine building to hit automotive latency and power budgets without sacrificing accuracy.

Integration on QNX/AUTOSAR Adaptive and real-time Linux, including scheduling, time synchronization, and safety-partitioned compute.

High-fidelity closed-loop simulation in CARLA with sensor models, traffic agents, and weather variation for perception and planning testing.

OpenSCENARIO/OpenDRIVE scenario libraries and parameter-sweep generation covering edge cases, cut-ins, and NCAP protocols.

HIL and SIL rigs that exercise ECUs and autonomy software against simulated sensors and vehicle dynamics for repeatable regression testing.

Physics-accurate digital twins of vehicle and sensors for behavior replay, what-if analysis, and continuous validation.

Replay of real-world drive logs through updated software to measure regressions and disengagement causes before OTA release.

Definition and dashboarding of disengagement rate, miss/false-positive rates, comfort, and time-to-collision KPIs across sim and fleet.

Test-track and public-road validation planning, safety-driver protocols, and data capture rigs for real-world verification.

High-throughput ingest, decoding, and storage of multi-terabyte camera/LiDAR/CAN drive logs with lineage and indexing.

Bounding-box, cuboid, polygon, and point-cloud annotation workflows accelerated by model-assisted auto-labeling and QA review.

Uncertainty- and disagreement-based data selection that surfaces rare and failure-prone scenarios for targeted labeling and retraining.

Searchable scenario and object databases with metadata tagging for balanced dataset construction and coverage analysis.

Automated training, evaluation, versioning, and deployment pipelines that close the loop from fleet data to validated model updates.

Hazard analysis, ASIL decomposition, safety-goal derivation, and safety-case development across the item lifecycle.

Safety-of-the-intended-function analysis addressing perception limitations, triggering conditions, and unknown-unsafe scenario reduction.

Automotive threat analysis (TARA), secure boot/OTA, and CSMS processes aligned to ISO/SAE 21434 and UNECE R155 regulation.

Vehicle-to-everything messaging for cooperative perception, intersection assist, and hazard warning over C-V2X and DSRC stacks.

Low-latency remote monitoring and remote-driving systems that let operators resolve edge cases the autonomy stack cannot handle.

Fleet orchestration, health telemetry, mission dispatch, and secure over-the-air update pipelines for autonomous vehicle fleets.