A
AcciVision AI
"Let's go Crash-less" & "No dash, No crash"
Created on 12th October 2025
•
A
AcciVision AI
"Let's go Crash-less" & "No dash, No crash"
The problem AcciVision AI solves
❗ The Problem It Solves
Road accidents are a leading cause of injury and death worldwide, and one of the biggest delays in saving lives is the time it takes to detect and report an accident. In many areas — especially in developing regions or on less-monitored roads — accidents go unnoticed for critical minutes or even hours.
Most existing traffic surveillance systems are passive, requiring human operators to monitor video feeds manually. This leads to:
- ⏱️ Delayed emergency response
- 👀 Missed incidents due to human fatigue
- 💸 Expensive infrastructure and monitoring costs
AcciVision-AI solves this problem by introducing an automated, real-time accident detection system that:
- Detects accidents from live video feeds using deep learning
- Logs snapshots and timestamps for review on a web dashboard
- Works on low-power edge devices, removing the need for centralized or cloud processing
This helps in:
- 🚑 Accelerating emergency response
- 🏙️ Supporting smart city infrastructure
- 🌍 Making traffic safety tech more accessible and affordable
- 👷 Assisting in automated traffic incident reporting for municipalities and highway operators
By turning existing video surveillance infrastructure into intelligent, automated systems, AcciVision-AI bridges the gap between observation and action.
Challenges I ran into
🚧 Problems We Ran Into
1. ⚙️ Model Accuracy & Overfitting
- Initially, the CNN-LSTM model was overfitting on the limited dataset, especially for non-accident scenarios.
- Balancing between real-time performance and model accuracy was a challenge — we had to carefully tune the architecture and apply dropout and regularization techniques.
2. 🎥 Real-Time Frame Processing
- Achieving low-latency video processing on edge devices (like Raspberry Pi) required optimizing how frames were captured, resized, and batched.
- Some inference delays were fixed by reducing input resolution and minimizing unnecessary preprocessing steps.
3. 🌐 Flask Dashboard + Live Stream Integration
- Integrating a live video feed with Flask and ensuring updated snapshots appeared in near real-time involved multiple refresh/render issues.
- Browser caching and manual image refresh hacks had to be introduced for consistent display.
4. 📂 Dataset Collection & Labeling
- Reliable accident footage for training was hard to find due to ethical and privacy issues.
- We had to rely on publicly available dashcam videos and manually labeled sequences, which was time-consuming and inconsistent.
5. 🧪 Testing on Edge Devices
- Deploying the project on a Raspberry Pi exposed several resource constraints:
- Memory usage spikes
- TensorFlow compatibility issues
- Slower FPS during inference
- Required model quantization and pruning to improve runtime performance.
Tracks Applied (1)
Open Track - Development
Discussion
Builders also viewed
See more projects on Devfolio