MAHE
Back to Tracks
Track_01 in Collaboration withHarman Automotive

AI IN
MOBILITY.

Reimagine the connected cockpit. Build advanced AI models for trajectory prediction, drivable space segmentation, and Bird's-Eye-View mapping.

Problem Statements

Problem Statement 1

Intent & Trajectory Prediction

FOCUS Behavioral AI & Temporal Modeling

PROBLEM STATEMENT In an L4 urban environment, reacting to where a pedestrian is isn't enough; the vehicle must predict where they will be. Participants must develop a model that predicts the future coordinates (next 3 seconds) of pedestrians and cyclists based on 2 seconds of past motion.

KEY FOCUS AREAS LSTMs/GRUs, Transformers, Social-Pooling layers, and Goal-conditioned prediction.

DATASET nuScenes. Access Here

Objectives
  • Process temporal sequence data (coordinates/velocity).
  • Account for "Social Context" (how pedestrians avoid each other).
  • Generate a multi-modal prediction (e.g., the 3 most likely paths).
Expected Outcomes & Metrics
  • Outcome: A model that inputs a history of (x, y) coordinates and outputs a sequence of predicted future (x, y) points.
  • ADE: Mean Euclidean distance between predicted and ground truth points.
  • FDE: Distance between the final predicted point and the actual final position.
Problem Statement 2

Real-time Drivable Space Segmentation

FOCUS Semantic Perception & Edge Cases

PROBLEM STATEMENT Level 4 vehicles must identify "Free Space"; areas where the car can physically move; regardless of whether lane markings exist. This track focuses on segmenting the road vs. everything else (curbs, construction barriers, sidewalks) in complex urban settings.

KEY FOCUS AREAS Encoder-Decoder architectures (U-Net, DeepLabV3+), Real-time backbones (MobileNet/EfficientNet), and Loss functions for class imbalance.

RESTRICTIONS Use of pre-trained models is strictly prohibited. Models must be trained from scratch.

DATASET nuScenes. Access Here

Objectives
  • Perform pixel-wise semantic segmentation of the drivable area.
  • Ensure high-frequency performance (inference speed is critical).
  • Handle "boundary" cases like road-to-grass transitions or water puddles.
Expected Outcomes & Metrics
  • Outcome: A real-time inference pipeline that outputs a binary or multi-class mask representing "Drivable" vs. "Non-Drivable."
  • mIoU: The primary accuracy metric for segmentation.
  • FPS & Architecture: Inference speed (FPS) is critical. Additionally, your model architecture and training epochs will also be evaluated.
Problem Statement 3

Bird’s-Eye-View (BEV) 2D Occupancy

FOCUS Geometric Transformation & Spatial Representation

PROBLEM STATEMENT Standard front-view cameras suffer from perspective distortion. For planning, L4 systems convert camera images into a 2D top-down "Occupancy Grid." Participants must transform front-facing camera features into a Bird's-Eye-View (BEV) map showing occupied (obstacles) and empty space.

KEY FOCUS AREAS Homography, Spatial Transformers, BEV-Former architectures, and Multi-View fusion.

DATASET nuScenes. Access Here

Objectives
  • Implement an "Image-to-BEV" transformation (e.g., Inverse Perspective Mapping or View Transformers).
  • Map 3D obstacles (cars, poles) onto a 2D ground plane.
  • Maintain spatial consistency (objects shouldn't "stretch" in the BEV view).
Expected Outcomes & Metrics
  • Outcome: A 2D grid/heatmap where each pixel represents a fixed area (e.g., 10cm x 10cm) and its probability of being occupied.
  • Occupancy IoU: How well the top-down occupancy matches the ground truth derived from LiDAR.
  • Distance-weighted Error: Errors closer to the ego-vehicle are penalized more heavily.

AI Track Submissions Closed

Submissions for the AI track are now closed. Thank you to all teams who submitted their work.

Submission Requirements
  • Working Code Required: We require the working code of your model from Round 1 itself. The provided PPT template must be fully completed, and your code must be hosted on a GitHub repository linked within the PPT. Evaluation will be strictly based on model metrics. Only 10 teams will be selected per problem statement.
  • Single Selection: Teams are permitted to select and submit a solution for exactly one problem statement.
  • Accessibility & Disqualification: Ensure all submission files are publicly accessible. If your presentation file cannot be opened or your GitHub repository is unviewable, your team will be disqualified immediately.
Submission window has ended