Traffic Light and Crosswalk Detection with MobileNetv2 SSD

This repository contains code and models for detecting traffic lights and crosswalks using MobileNetv2 SSD (Single Shot MultiBox Detector).

Prerequisites

Make sure you have the following dependencies installed:

• Python (3.6 or higher)
• TensorFlow (2.8)
• Roboflow

Dataset Information

The dataset used for training and testing traffic lights and crosswalks detection model consists of 3,793 images.

Class	Count
Traffic Light (Red)	1182
Traffic Light (Green)	1138
Traffic Light (Yellow)	827
Crosswalk	1266
Null Examples	241

Download

Model Evaluation

The trained traffic light and crosswalk detection models were evaluated using the following configuration:

Model	Batch Size	Epochs	mAP@50%
ssd-mobilenet-v2-fpnlite-320	16	45000	90,56%

• Model: MobileNetv2 SSD (Single Shot MultiBox Detector) was used for object detection.
• Batch Size: The batch size during training and evaluation was set to 16.
• Epochs: The models were trained for 45000 epochs.
• mAP@50%: The mean Average Precision (mAP) was calculated at the intersection over union (IoU) threshold of 50%.

The average mAP across different intersection over union for the trained models is 58,82%. The mAP is a metric that measures the accuracy of object detection models.

Loss Graph

mAP Graph

Testing Result

The following images show the result of traffic light and crosswalk detection.

Red	Yellow	Green	Crosswalk