Discernible Effect of Video Quality for Distorted Vehicle Detection using Deep Neural Networks

In the transport sector, the rapid development of applications in communications and information technologies has led to the evolution of the intelligent transportation system (ITS). For sustainable development of road transport and traffic management, ITS played a vital role, especially in traffic efficiency and safety. However, with better road infrastructure, the growing number of vehicles on the road, and fast-track highways, vehicle detection in real-time is an important and challenging task. The real-time vehicle detection systems must distinguish and find vehicles in foggy, cloudy, and rainy environments, as these environmental conditions may reflect the noise effect in the captured videos. Therefore, in this work, the effect of video quality on object detection in self-driving scenarios is analyzed and tested. Deep learning algorithms, Res Net and YOLO, trained with high-quality video streams, are used for object detection in lower quality video streams. In this regard, the video quality is reduced with the increase in the quantization parameter using H.264, by the addition of Gaussian noise, and by introducing motion blur. Following this, the encoded videos are processed for vehicle detection, and it is found that with an increase in quantization parameters the effect on vehicle detection is negligible. In the degraded video through the Gaussian noise, the vehicles' detection fails above 1.5 dB added noise, while in the case of motion blur the detection failed when the motion blur exceeded 53 %.