After clicking “Watch Now” you will be prompted to login or join.
Object Recognition and Tracking Utilizing Millimeter-Wave Radar by Deep Neural Networks
Tokihiko Akita, Toyota Technological Institute
GTC 2020
All-weather sensors are necessary for autonomous-driving Level 3 and higher. Millimeter-wave radar is the most robust sensor for adverse weather. However, the signal is noisy and fluctuated, and the resolution is low. Thus, recognition using the radar is difficult. Deep-learning algorithms are an effective solution. We'll show a method to classify and track objects in driving scenes with a high-resolution millimeter-wave radar applying long short-term memory. We designed and compared various types of input features and LSTM for our measured dataset and achieved high accuracy through cross validation. We'll also show a method to reconstruct shapes of parking spaces and cars with convolutional neural networks. Parking cars were scanned with side radar. The reflection signals were accumulated, and the shape was estimated by semantic segmentation framework, applying CNN for the ground-truth shape, annotated by a lidar.