A Low Cost and Real Time Vehicle Detection Using Enhanced YOLOv4-Tiny

Just as the accuracy of the detectors is important, so the speed of their performance is no less important, especially in applications that require real-time response processing. In addition to real-time processing, vehicle traffic systems require detection models to be lightweight and deployable to peripheral devices, which typically have limited processing and memory capacities. In this paper, we propose an enhancement to the YOLOv4-tiny detector in order to reduce memory usage and increase image processing speed. Our key improvement is on the detector backbone so that less floating-point operations are generated while still maintaining vehicle feature information, thus that detection accuracy is not compromised. On a vehicles dataset, we trained and tested the proposed model, then compared the results to the most common models in the same category. The experimental results revealed that our proposed model is 7.9% more FPS and 0.5% less memory consumption than the base model. Our proposed model outperforms the basic model in terms of speed and memory consumption, while maintaining a level of accuracy that is nearly equal to the basic model.