Dynamic Speed Estimation of Moving Objects from Camera Data

Recent advances in computer vision techniques have led to increasing interest in the use of vision as the only mechanism to estimate the vehicle speed in certain applications. The challenge of making accurate estimations of distances and speeds arises from the discrete nature of video sensors that project the 3D world into a 2D discrete plane. This work will study the concept of estimating the speed of an object by only using the camera data. Rather than using a separate mechanism dependent on camera specifications, we consider a convolutional neural network fed recurrent neural network to estimate the speed at each time step. The video data can be fed to an object detection algorithm to extract the features and patterns in the videos. The network architecture is designed to make use of pre-trained object detection models to extract features from camera video feed and embed them into a data cluster. This cluster can represent the spatiotemporal information of the objects in the video frames. A recurrent neural network is then used to learn the association of the motion speed values with respect to the changes in the features within the cluster. Based on this structure, both the object detector and the recurrent neural network can be trained end-to-end to detect and estimate the speed of objects in videos. This network architecture would not require additional sensors for calibration, while providing a cost-effective solution for mobile robotic applications such as autonomous ground vehicles or drones. The proposed methodology, network architecture and its components will be discussed in detail. This paper also presents a video data case study and preliminary results to demonstrate the development and use of the speed estimation network. Some future work is planned to improve the algorithm accuracy and computing efficiency of the proposed architecture.