Embedded and real-time vehicle detection system for challenging on-road scenes

摘要

Vehicle detection is an important topic for advanced driver-assistance systems. This paper proposes an adaptive approach for an embedded system by focusing on monocular vehicle detection in real time, also aiming at being accurate under challenging conditions. Scene classification is accomplished by using a simplified convolution neural network with hypothesis generation by SoftMax regression. The output is consequently taken into account to optimize detection parameters for hypothesis generation and testing. Thus, we offer a sample-reorganization mechanism to improve the performance of vehicle hypothesis verification. A hypothesis leap mechanism is in use to improve the operating efficiency of the on-board system. A practical on-road test is employed to verify vehicle detection (i.e., accuracy) and also the performance of the designed on-board system regarding speed. driver assistance systems; embedded systems; monocular vision; scene recognition; vehicle detection