Abstract
The growing development of sensing and deep learning technologies greatly extend vehicle perception capabilities, which are critical for intelligent vehicles. However, to achieve high quality perception within the vehicle's onboard electrical control unit (ECU), the high efficiency of edge computing and data encoding is essential. Unfortunately, this is a less studied upon topic from the deep learning perspective, as it is always associated with high computational overhead and data throughput, which limit the wide applications of deep learning in intelligent connected vehicles. To solve this problem, we propose a novel deep-learning architecture to simultaneously implement vehicle on-board sensing and detected vehicle identification (ID) creation. To this end, a single pass deep neural network with triplet structure and hashing encoder are delicately designed. The remarkable properties of the vehicle ID include discriminant representation, privacy preservation, geolocation awareness, and minimal communication requirements. All these potentials make the proposed vehicle ID be a viable representation that can meet the strict requirements of realtime large-scale sensing analytics in practical applications of intelligent vehicle systems.
Published Version
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call