Developing an Advanced Deep Learning SSD Algorithm Using Computer Vision Approaches to Enhance Vehicle Detection

Abstract

The development of Computer vision and Deep learning frameworks plays a crucial role in Autonomous driving and real-time intelligent traffic surveillance. The Single Shot Detector (SSD) architecture is well known for its ability to perform inference quickly, and generically configured anchors are problematic because of the severe localization errors they cause. This article proposes an optimized architecture for SSD-based vehicle detection that addresses these challenges in two main steps: (1) using an optimized edge-preserving median filter to remove semantic noise in the image caused by the sensors while preserving the sharpness of the image boundaries and (2) applying custom-scaled anchor box aspect ratios specifically designed to capture the real size of the vehicles on the road. Applied to the standard KITTI Vision Benchmark dataset, the proposed framework's Mean Average Precision (mAP) and recall were 95.90% and 96.60%, respectively. The proposed model is compared with the baseline networks and is shown to significantly outperform them, achieving mAP 89.50% for vanilla SSD300 and 94.10% for YOLOv5. The proposed method reached high reliability with a balanced precision-recall rate that is highly suitable for real-time traffic monitoring systems.