Speed estimation using deep learning with optical flow

For automatic driving, it is necessary to perform a lot of simulations that consider various kinds of situations including hazardous events. In the simulations, driving scenarios are needed and should be generated from real driving records. On the other hand, most vehicles have dashcams, and the video taken by the camera is used as evidence in case of car accidents or near-miss reports. In addition, Artificial intelligence (AI), which is achieved by using deep learning methods, is widely used to recognize many kinds of unknown objects. This paper reports a method and the result of deep learning-based estimation of automobile speed. We employ 3D CNN as the basic learning model and four-fold cross-validation as the accurate measure. First, 16-frame original images obtained by a dashcam, whose resolution is reduced for faster learning, are used for learning and validation. Four-fold cross-validation and estimation results are reported for the speed range from 70 [km/h] to 100 [km/h] in every 5 [km/h]. Then, the images with optical flow generated by two types of methods, which are the Lucas-Kanade and Gunner-Farnebäck methods, are utilized, and the estimation errors are reported for the speed range from 70 [km/h] to 100 [km/h] in every 1 [km/h]. We have confirmed that optical flow is useful for estimating automobile speed from dashcam video.