Object Detection on Road: Vehicle's Detection Based on Re-Training Models on NVIDIA-Jetson Platform.

Area of Science:

• Computer Vision and Artificial Intelligence
• Embedded Systems Engineering
• Transportation and Traffic Management

Background:

• Growing use of AI and deep learning (DL) for vehicle classification and detection on embedded devices.
• Deployment constraints on embedded systems include computational cost and response time, especially in congested urban areas.
• Need to balance model accuracy, embedded system type, and dataset characteristics for urban traffic analysis.

Purpose of the Study:

• To evaluate the trade-off between model accuracy and computational load for vehicle detection in congested urban environments.
• To identify the optimal SSD-based deep learning model for implementation in ADAS embedded systems.
• To assess the impact of data augmentation on detecting minority vehicle classes.

Main Methods:

• Acquisition and manual labeling of urban traffic videos from Lima, Peru.
• Training of three Single Shot Detector (SSD)-based models (MobileNetV1-SSD, MobileNetV2-SSD-Lite, VGG16-SSD) on an NVIDIA Jetson Orin NX platform.
• Utilized a methodology adapted from the CRISP-DM approach, including data augmentation via contrast adjustment.

Main Results:

• VGG16-SSD achieved the highest mean average precision (mAP ≈90.7%) but required longer training.
• MobileNetV1-SSD (512×512) demonstrated comparable performance (mAP ≈90.4%) with significantly reduced training time.
• Data augmentation improved detection rates for minority classes like Tuk-tuks and motorcycles.

Conclusions:

• MobileNetV1-SSD (512×512) offers the best compromise between accuracy and computational efficiency for ADAS in congested urban settings.
• The study highlights the importance of selecting appropriate AI models based on specific deployment constraints and traffic scenarios.
• Contrast adjustment data augmentation is effective for enhancing the detection of underrepresented vehicle types.