See What You Can't.

An open-source AI copilot that brings Advanced Driver Assistance Systems to bicycles. Powered by computer vision and real-time tracking algorithms.

HOW IT WORKS

TALOS is an open-source blind-spot monitoring system built to protect cyclists in urban environments. Running on a Raspberry Pi 5 and Hailo NPU, this system utilizes a custom YOLOv8 computer vision pipeline to detect vehicles and warn riders of approaching threats in real-time. I engineered a custom dataset using class merging and oversampling techniques. This fine-tuning process resulted in a 41% increase in recall for minority classes (like motorcycles) compared to stock models.

Test TALOS!

// OPERATIONAL_CAPABILITIES

SYSTEM ARCHITECTURE

01

Global Shutter Optics

Standard cameras warp high-speed objects. Talos utilizes the IMX296 Global Shutter Sensor to capture distortion-free, mathematically accurate imagery of moving vehicles, ensuring precise detection at any relative velocity.

02

Real-Time Edge AI

Powered by the Hailo-8L NPU. All neural network inference happens locally on the device with ~13ms latency. No cloud dependencies, no lag, and complete privacy by design.

03

Looming Vector Analysis

Goes beyond simple object detection. Talos uses Optical Expansion and Vector Calculus to calculate the rate of expansion (dA/dt) of approaching objects, predicting Time-to-Collision (TTC) with high precision.

04

Intelligent Threat Filtering

Silence is golden. The system distinguishes between a car passing safely and a vehicle on a collision course, filtering out false positives to provide audible warnings only when a genuine threat is detected.

05

Day & Night Mastery

Fine-tuned on the Valdiolus Rear View Camera dataset. Talos is trained to differentiate between cars, trucks/buses, motorcycles, cyclists, and pedestrians across varied lighting conditions, from high-noon glare to low-light dusk.

06

Open Source Architecture

Built on Python and YOLOv8. Talos is fully modular and open-source, allowing engineers and developers to audit the code, customize detection models, and extend hardware capabilities.

// OPERATIONAL_LOGIC

DETECTION LOOP

PHASE 01

VISUAL ACQUISITION

IMX296 sensor utilizes Global Shutter technology to capture blur-free frames in high-vibration environments. Images are buffered directly to system RAM to minimize I/O latency before neural inference.

PHASE 01

VISUAL ACQUISITION

PHASE 01

VISUAL ACQUISITION

PHASE 02

NEURAL INFERENCE

A custom version of the YOLOv8 neural network (NCNN) classifies approaching threats including vehicles, cyclists, and pedestrians. The pipeline utilizes hardware-accelerated math to maintain real-time tracking accuracy without overheating the CPU.

PHASE 02

NEURAL INFERENCE

PHASE 02

NEURAL INFERENCE

PHASE 03

THREAT ASSESSMENT

Looming detection algorithms calculate the expansion rate of the target. If Time-to-Collision drops below 2.5s, the system triggers visual warnings. Users can stream the camera view to their phone.

PHASE 03

THREAT ASSESSMENT

Looming detection algorithms calculate the expansion rate of the target. If Time-to-Collision drops below 2.5s, the system triggers visual warnings. Users can stream the camera view to their phone.

PHASE 03

THREAT ASSESSMENT

Looming detection algorithms calculate the expansion rate of the target. If Time-to-Collision drops below 2.5s, the system triggers visual warnings. Users can stream the camera view to their phone.