Training Data

Overview

The CycleSafely system relies on high-quality training data collected from real-world traffic scenarios and benchmark datasets. This page describes the data collection process, dataset characteristics, and model training strategies used to develop the collision detection pipeline.

Data Collection Setup

Sensor Configuration

Data is collected using a multi-sensor setup mounted on bicycle handlebars:

• RGB Cameras: High-resolution video for vehicle detection and ground truth annotation
• Depth Sensors: Kinect or depth-from-stereo for distance measurements
• LIDAR: Livox Mid-360 for accurate 3D point clouds
• GPS + IMU: Position and orientation tracking for registration

Recording Pipeline

The recording system captures synchronized multi-modal data:

• Timestamped RGB video streams
• Registered point clouds from LIDAR
• GPS coordinates and IMU measurements
• Depth images (captured or estimated using ML)
• Metadata (weather, lighting conditions, traffic density)

Dataset Characteristics

Mobile Optimization Techniques

1. Model Compression

Optimizations for deployment on smartphones and tablets:

• Quantization: INT8 quantization for faster inference
• Pruning: Removing redundant network parameters
• Knowledge Distillation: Training smaller models from larger teachers
• Architecture Search: Lightweight architectures (MobileNet, EfficientNet)

2. Efficient Processing

• Frame skipping and adaptive resolution
• Region-of-interest focused processing
• Temporal coherence exploitation
• GPU/NPU acceleration on mobile devices

3. Data Pipeline Optimization

• Streaming data processing
• Asynchronous sensor fusion
• Buffered recording with compression
• Selective high-resolution capture

Datasets Used

Primary Datasets

• KITTI: Real-world LiDAR and camera data from Karlsruhe, Germany
• CARLA: Synthetic data from the autonomous driving simulator
• nuScenes: Multi-modal dataset with diverse urban scenarios
• Argoverse: High-definition maps and trajectory data
• Waymo Open Dataset: Large-scale autonomous driving data

Model Training

Object Detection (VoxSeT):

• Pre-trained on KITTI 3D object detection benchmark
• Direct point cloud processing without BEV conversion
• Trained to detect cars, cyclists, and pedestrians
• Optimized for real-time inference on GPU

Trajectory Prediction:

• Motion-based algorithm using velocity and acceleration
• Curve fitting with polynomial and clothoid methods
• Optimized parameters through systematic evaluation
• Independent of environmental point cloud data