Performance Comparison of Faster R-CNN and SSD in Vehicle Detection from Aerial Traffic Videos

Vehicle detection is essential for urban planning, traffic management, and various security and surveillance applications. This study compares the performance of two popular deep learning algorithms, Faster R-CNN and SSD, for vehicle detection using traffic video recordings obtained from unmanned aerial vehicles. The dataset, comprising images taken from diverse angles, heights, and lighting conditions, was utilized to assess the effectiveness of both models. Faster R-CNN demonstrated superior accuracy in detecting vehicles and motorcycles under various conditions, including night scenarios and scenes with obstacles such as shadows, traffic signs, and light poles. However, it struggled with detecting very small objects and those with color tones similar to the background. In contrast, while SSD performed well under optimal lighting, it exhibited limitations in detecting smaller vehicles partially obscured by environmental elements. Quantitative analysis revealed that Faster R-CNN had 13.21% higher precision, 12.5% higher recall, and 10.04% higher mAP compared to SSD. Despite its superior detection performance, Faster R-CNN's longer detection time indicates that SSD may be more suitable for real-time applications where speed is crucial.