AI Perception and Planning for Autonomous Driving

AI Perception and Planning for Autonomous Driving

Perception + Planning — it's the combination that turns sensor data flow into vehicle control commands. We, a team of engineers with 10+ years in Computer Vision and Robotics, solve this problem systemically: from sensor calibration to deployment on the onboard computer. With over 10 years of experience and 50+ successful projects, we deliver robust perception systems. The challenge of domain gap between simulation and reality is the main difficulty: even with mAP >0.9 on benchmark datasets, the system may lose an object on wet asphalt or incorrectly estimate a pedestrian's trajectory. Domain randomization, described in Wikipedia article on domain randomization, is a key technique. Without its application, accuracy drops 15–25% on real data, and fixing errors at the validation stage costs hundreds of thousands of dollars.

Methods to Bridge Simulation-to-Reality Gap

Domain randomization — key technique: we randomly vary textures, lighting, and weather in the simulator (CARLA, SUMO). Real2Sim — transfer real scenes to virtual environment via NeRF. Curriculum learning: first simple scenes, then corner cases. Without this, the model loses 15–25% mAP on real data. According to our estimates, proper application of domain randomization reduces errors by 25%, saving a significant amount at the validation stage.

Why Correct Detection Model Selection Matters

BEVFusion outperforms SECOND by 8 points in mAP but requires 3x more computational resources. For the onboard NVIDIA Drive Orin (128 TOPS), we use TensorRT-optimized BEVFusion at 100ms, achieving 2x speedup over standard implementation. In complex scenarios, we escalate to UniAD with reduced FPS. We validate models on open datasets such as nuScenes.

How We Calibrate Sensors

Calibration is the first and critical step. We use target-based method for LiDAR-camera fusion: set up a chessboard, collect 3D point-to-pixel correspondences, solve Perspective-n-Point problem. Accuracy — up to 0.1° in angle and 1 cm in translation. Calibration is repeated after every sensor disassembly.

Sensor Stack and Fusion

Autonomous systems of level L3+ work with multiple sensor types simultaneously:

import numpy as np
import torch
from mmdet3d.models import build_detector
from mmdet3d.apis import inference_detector

class PerceptionPipeline:
    def __init__(self, config: dict):
        # 3D detector: BEVFusion or SECOND
        self.detector_3d = build_detector(config['detector_cfg'])
        self.detector_3d.load_checkpoint(config['checkpoint'])

        # 2D camera detector: YOLOv8 or DETR
        self.cam_detector = torch.hub.load('ultralytics/ultralytics',
                                            'yolov8l', pretrained=True)

        # LiDAR → camera projection matrices
        self.lidar2cam = np.array(config['lidar2cam_matrix'])
        self.camera_intrinsics = np.array(config['cam_intrinsics'])

    def fuse_lidar_camera(self, point_cloud: np.ndarray,
                           images: list[np.ndarray]) -> dict:
        """
        LiDAR gives accurate 3D coordinates and range,
        camera gives semantics (object type, traffic light color).
        BEVFusion combines into a single Bird's Eye View representation.
        """
        bev_features = self._to_bev(point_cloud)
        cam_features = [self.cam_detector(img) for img in images]

        # Project LiDAR points onto camera plane
        pts_3d_cam = self._project_lidar_to_cam(point_cloud)

        return {
            'bev_features': bev_features,
            'cam_detections': cam_features,
            'projected_points': pts_3d_cam
        }

Planning: From Perception to Trajectory

class MotionPlanner:
    def __init__(self, config: dict):
        self.dt = 0.1  # time step 100ms
        self.horizon = 5.0  # planning horizon 5 sec
        self.safety_margin = 0.8  # meters

    def plan_trajectory(self, ego_state: dict,
                         detected_objects: list[dict],
                         hd_map: dict) -> np.ndarray:
        """
        IDM (Intelligent Driver Model) + potential fields.
        For complex scenarios: RL or transformer (PDM-Closed).
        """
        # Candidate trajectories from generator
        candidates = self._generate_candidates(ego_state)

        # Assess safety of each trajectory
        scores = []
        for traj in candidates:
            collision_risk = self._collision_check(traj, detected_objects)
            lane_keep = self._lane_keep_cost(traj, hd_map)
            comfort = self._comfort_cost(traj)

            total_cost = (3.0 * collision_risk +
                          1.5 * lane_keep +
                          0.5 * comfort)
            scores.append(total_cost)

        best_idx = np.argmin(scores)
        return candidates[best_idx]

    def _collision_check(self, trajectory: np.ndarray,
                          objects: list[dict]) -> float:
        """TTC (Time-To-Collision) for each object"""
        min_ttc = float('inf')
        for obj in objects:
            ttc = self._compute_ttc(trajectory, obj)
            min_ttc = min(min_ttc, ttc)

        # TTC < 2 sec = high risk
        return 1.0 / max(min_ttc, 0.1)

Work Process

1. Analytics and calibration — on-site visit, data collection from the vehicle, sensor calibration (LiDAR-camera).
2. Architecture design — model selection, define fusion pipeline, set latency and precision requirements.
3. Pipeline implementation — write detection, tracking, prediction, and planning code. Integrate with simulator.
4. Testing — A/B tests on open datasets (nuScenes, Waymo), validation on collected data, stress tests for corner cases.
5. Onboard deployment — optimization via TensorRT, ONNX Runtime, inference on target platform (NVIDIA Orin/Thor).

Approximate Timelines

Cost is calculated individually after analyzing your data and requirements. Typical budgets start at $50,000 for L2 ADAS.

What's Included

• Sensor calibration and reference data collection
• Development of perception pipeline (detection, tracking, prediction)
• Model training and adaptation for your scenarios
• Integration with planning and control
• Architecture and API documentation
• Training your team to work with the system

Common Mistakes and Our Experience

A frequent issue is overfitting to a specific scenario. We guarantee generalization through domain randomization and tests on independent data. Our certified specialists have 50+ completed projects in autonomous driving. For example, a recent L3 system based on BEVFusion and IDM planner: in 18 months achieved mAP 70.5 on nuScenes with 110ms latency on Orin, and calibration automation reduced time by 40%.

Get a commercial proposal with a detailed work plan — contact us to evaluate your project. We will offer a turnkey solution considering your timeline and budget. Order a consultation to discuss details.

How Distribution Shift Kills CV Model Metrics in Industry

On a production line, a camera is installed to control product quality. The model is trained on 10,000 labeled images—test accuracy mAP 0.84. Deployed to production, and in the first week it misses 30% of defects. Lighting on the line changes between shifts; distribution shift nullifies the metrics. This is a classic story with computer vision in industry, where pattern recognition fails without proper drift handling.

Our engineers, with experience from 60+ computer vision projects, know how to eliminate such scenarios. We guarantee stable model performance under real conditions.

Object Detection: YOLO, RT-DETR, and Everything in Between

YOLO is the standard for real-time detection. YOLOv8 and YOLOv11 from Ultralytics are the most used versions in production: simple API, active community, built-in validation, and export to ONNX/TensorRT. For tasks with high accuracy requirements and less critical latency, RT-DETR, a transformer-based architecture without NMS, gives better mAP on COCO at comparable speed to YOLOv8l.

A typical mistake when training a detector: dataset of 8000 images, 3 classes, fine-tune YOLOv8m—F1 0.73 on validation. Look at confusion matrix—one class is almost never detected. Cause: imbalance 1:23. Solution: oversampling rare class, focal loss for objectness, augmentations (Mosaic, MixUp disabled for rare class as they "blur" it). Transfer learning is mandatory: pretrained on COCO weights reduces data requirement by 10 times. Fine-tuning on 500–2000 domain images yields a working model in 1–2 days on a single GPU.

For edge deployment: export to ONNX → TensorRT engine. YOLOv8n in TensorRT FP16 on Jetson AGX Orin gives 150+ FPS at P99 latency < 8 ms—3 times faster than ONNX Runtime without TensorRT. On server A10G: 700+ FPS for YOLOv8n in TensorRT INT8.

How Does Fine-Tuning YOLO Help in Pattern Recognition?

Suppose you need to find micro-defects on a metal surface—a task with high resolution and class imbalance. We use YOLOv8m pretrained on COCO and fine-tune on 2000 proprietary images. Apply augmentations Mosaic, MixUp, random perspective. After 200 epochs, mAP 0.5 reaches 0.93. Key techniques:

• Focal loss for the objectness head—reduces contribution of easily classified examples.
• Class-balanced sampling—equalizes representation of rare classes.
• Test Time Augmentation (TTA)—increases recall by 5–7% through averaging over flips and scales.

Get a consultation on architecture selection for your task—contact us.

Segmentation: SAM, Mask R-CNN, and Instance Segmentation

SAM (Segment Anything Model) from Meta changed the approach to segmentation. SAM 2 works with video, supports object tracking across frames—for interactive object selection by point or bbox, it's the best out-of-the-box choice. For production instance segmentation without interactive prompting, Mask R-CNN or YOLOv8-seg are used. YOLOv8-seg trains like a regular detector with additional masks, convenient in the same pipelines. Semantic segmentation (each pixel is a class) uses SegFormer, DeepLabV3+. SegFormer-B5 provides a good balance of accuracy and speed for satellite imagery or medical segmentation.

Case study: cell segmentation on microscopic images. Dataset of 400 images with manual annotation. Training Mask R-CNN on ResNet-50 backbone gave IoU 0.61—poor. Problem: objects (cells) overlap; standard NMS kills overlapping predictions. Solution: switch to cellpose (specialized architecture for biomedical tasks) + soft-NMS. IoU increased to 0.79.

OCR: When Tesseract Fails

Tesseract is a starting point for simple tasks: printed text, good lighting, straight layout. As soon as there are handwritten elements, non-standard fonts, perspective distortions, or multi-column layouts, Tesseract degrades quickly.

PaddleOCR is a production-grade solution: text block detection + recognition + structural analysis. Works out of the box for 80+ languages, including Russian. Supports tables and complex document structures. TrOCR (Microsoft) is a transformer OCR with strong results on handwritten text. For Russian handwritten text, fine-tuning is needed: the base model is trained mostly on Latin script.

What to Do When Tesseract Cannot Handle Pattern Recognition on Documents?

For tasks like "extract data from invoices/contracts/passports," we use LayoutLMv3 or Donut—these models understand document layout, not just text. Integration via Hugging Face Transformers, fine-tuning on 200–500 annotated documents. Typical pipeline:

1. Preprocessing: deskew, denoising, binarization via OpenCV.
2. Text block detection: PaddleOCR detection or CRAFT.
3. Recognition: PaddleOCR recognition or TrOCR.
4. Post-processing: normalization, validation via regex or LLM for structured fields.

For documents with fixed structure, template matching + OCR by coordinates is often more reliable than an end-to-end solution.

Face Recognition: Identification and Verification

Face recognition = detection + alignment + embedding + matching. Each stage matters.

Detection: RetinaFace or InsightFace for accurate face localization and keypoints. MTCNN is older but reliable. Embedding: ArcFace (InsightFace) is state-of-the-art for face recognition embeddings. Models iresnet50/iresnet100 pretrained on MS1MV3 (5M identities). Embedding vector 512 float32, comparison by cosine similarity. Threshold tuning: decision threshold is a critical parameter. At threshold 0.6, typical FPR on LFW benchmark is 0.001, TPR is 0.985. In production, threshold must be calibrated to the real distribution: people in masks, with changed appearance, different lighting conditions. Liveness detection is mandatory: MiniFASNet—lightweight model on CPU; FaceX-Zoo contains several pretrained liveness detectors.

Video Analytics

Video is a sequence of frames plus a temporal dimension. A naive approach—detecting on every frame—is expensive.

Tracking: ByteTrack and BoT-SORT are the standard for multi-object tracking. They work on top of any detector, adding persistent IDs to objects across frames—enabling object counting, motion tracking, velocity.

Optimization: not every frame needs processing. For static scenes, detect every 5–10 frames, with tracking in between. For event detection (person entering a zone), background subtraction (OpenCV MOG2) serves as a lightweight pre-filter before neural detection. Action recognition: SlowFast, VideoMAE for action classification. Heavy models—for production use ONNX export + TensorRT or offline processing.

How to Measure Pattern Recognition Model Quality in Production?

Quality monitoring is key to MLOps. We track:

• Prediction confidence distribution.
• Share of low-confidence predictions (indicator of OOD data).
• Drift of input images via feature distribution (embeddings from backbone).

A drop in average confidence from 0.87 to 0.71 over a week is an early signal of distribution shift. NVIDIA Triton Inference Server recommends tracking these metrics via Prometheus. Our certified engineers set up monitoring and guarantee SLA for inference quality.

Deployment of CV Models

For online inference, we use Triton Inference Server (NVIDIA)—production standard for serving CV models. Supports TensorRT, ONNX, PyTorch, dynamic batching, multiple instances. REST and gRPC API. We guarantee stable operation under load.

Edge deployment: ONNX Runtime on ARM/x86 CPU. TensorFlow Lite for mobile devices. OpenVINO for Intel CPU/GPU/VPU—gives 2–3× speedup on Intel hardware compared to ONNX Runtime. After deployment, we hand over the model with documentation and train personnel.

What Is Included in the Work

Deadlines and Cost

A prototype detector on existing data takes 1–2 weeks. Production system with optimization for target hardware takes 4–8 weeks. Full cycle including data labeling (1000–5000 images) takes 2–4 months. Cost is calculated individually for each task. Typical savings from implementing a quality control system can be significant per production line.

We have been in the market for over 5 years and completed 60+ computer vision projects. We will evaluate your project end-to-end—request a consultation to get a quote and technical proposal.