Deep Learning-Based Framework for Traffic Estimation for the MLK Smart Corridor in Downtown Chattanooga, TN

In this paper we introduced a deep learning-based framework for vehicles detection, tracking and speed estimation. YOLOv7 has been trained to detect and classify vehicles into Sedan, SUV, Pickup truck and Bus with a reported MAP of 0.69. For vehicles re-identification, we investigated the use of DeepSort with a Siamese network as its deep feature extractor (revision of DeepSort). We trained both the proposed Siamese network and the CNN of the DeepSort on the UA-DETRAC dataset, then tested them on the KITTI dataset and the revised model showed an average reduction rate of 71% in the IDSW rate when compared to DeepSort. Prior to the speed estimation we calibrated a total of 11 cameras assuming a pinhole model to estimate the parameters of each of the cameras, then through perspective transformation, reference object scaling and zones creation, we were able to estimate vehicles' speeds with an error rate of 0.516 mph. The amount of rich and accurate data our framework produce paves a way for different potential application, such as travel time estimation for multi-camera tracking and creating speed data benchmarks.