Vehicle Detection in Urban Traffic Surveillance Images Based on Convolutional Neural Networks with Feature...

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Urban traffic surveillance requires real-time vehicle detection and classification.
• Accurate localization of small vehicles in complex scenes is challenging.
• Existing methods like Single Shot MultiBox Detector (SSD) face limitations in real-time performance and accuracy.

Purpose of the Study:

• To develop an improved vehicle detection framework for real-time urban traffic surveillance.
• To enhance the performance of the Single Shot MultiBox Detector (SSD) for vehicle detection.
• To achieve accurate localization and classification of vehicles in video sequences.

Main Methods:

• Proposed a novel framework named DP-SSD, enhancing the conventional Single Shot MultiBox Detector (SSD).
• Utilized different feature extractors for localization and classification within a single network.
• Incorporated deconvolution (D) and pooling (P) between layers in the feature pyramid to enhance feature extractors.
• Adjusted the scale of default boxes to improve the training guidance for smaller vehicles.

Main Results:

• DP-SSD demonstrated efficient real-time vehicle detection on UA-DETRAC and KITTI datasets.
• Achieved 75.43% mean average precision (mAP) at 50.47 frames per second (FPS) with a 300x300 input size on the UA-DETRAC test set.
• Reached 77.94% mAP at 25.12 FPS with a 512x512 input size on the UA-DETRAC test set.
• Showed comparable accuracy to state-of-the-art models, outperforming most except YOLOv3.

Conclusions:

• DP-SSD offers an effective solution for real-time vehicle detection in urban traffic surveillance.
• The proposed method achieves a strong balance between accuracy and speed.
• DP-SSD provides a viable alternative for intelligent transportation systems and traffic management.