In our previous project funded by BDD, we proposed SqueezeDet, a convolutional neural network (CNN) for image-based object detection that achieved remarkable latency and energy reduction compared with previous works. Therefore, we would like to explore efficient architectures to model spatial-temporal evolution of videos. Autonomous driving requires the perception model to achieve real-time inference speed (>25 frames per second) and low power consumption (~20W).
Tracking can provide historical trajectories of other traffic objects, which are important for behavior prediction for other objects and essentially trajectory planning for the ego vehicle.Ĭ) Efficient neural network architectures for spatial-temporal modelling. We plan to explore the following three aspects:Ī) From static image to videos: with video object detection we can better utilize spatial-temporal information of videos to perform more accurate, consistent and robust detection on each frame and potentially save redundant computations.ī) From detection to tracking. In contrast, what is needed for autonomous driving applications are models that are able to perform accurate detection and tracking on video sequences in real-time, and to perform these computations in an energy efficient manner. In recent years, we have witnessed remarkable increases in accuracy of object detection through the use of Deep Neural Networks. However, previous efforts have been primarily focusing on: a) improving accuracy measured by bounding box overlapping (intersection-over-union), and b) detection on static images. Object detection and tracking are fundamental problems in the computer vision community and are fundations for autonomous driving perception.
0 kommentar(er)
