Transfer Learning in ML-based Radar Systems for Automotive Applications

Frequency modulated continuous wave radar systems are a vital and widespread component in modern vehicles. Their ability to sense the surrounding environment quickly and reliably is heavily used in today's driving assistance systems. Latest applied research introduced machine-learning based algorithms for real time detection and classification of radar targets without the need of traditional radar signal processing. The reliable and fast characteristics of these systems are a huge improvement of protecting vulnerable road users like pedestrians and bicyclists, overcoming the traditional methods of radar signal processing. In exchange, machine learning based systems need to be trained with a suitable dataset representing as many as possible real-world case scenarios. Refining adequate datasets from raw sensor data is a time-and cost-consuming effort. In order to achieve the best results, the sensor data must be manually processed and labeled after measuring the use-case in a real-world environment. Hence the application of transfer learning approaches derived from computer-based vision algorithms in automobile radar processing is investigated. A synthetical dataset including labeled radar readings is proposed to eliminate the necessity for manual dataset construction. Transfer learning approaches for machine-learning based radar processing algorithms are explored using a synthetic dataset to enhance detection and classification performance in automotive-related applications. The use of transfer learning already improves the effectiveness of activities based on machine learning for detection and classification in related applications like ship detection with synthetic aperture radar.