Vehicular Crowdsensing (VCS) aims to collect sensing data over a range of areas using a large number of on-board sensors and resources in intelligent vehicles. The mobility of vehicles allows for large-scale mobile sensing data, but it remains a challenging problem to recruit the right participating vehicles and to actively maximize the sensing benefits. In this paper, we formulate the vehicle recruitment problem as the maximizing completion rate with limited budget problem (MCRLB) and prove that it is NP-complete. A hybrid recruitment scheme based on deep learning in vehicular crowdsensing (HR-DLVCS) is proposed in this paper, which consists of two phases: an opportunistic vehicle recruitment phase and a participatory vehicle recruitment phase. In the first phase, a deep learning-based opportunistic vehicle recruitment algorithm (DL-OVR) is proposed to maximize the sensing task completion rate within a limited budget. It aims to recruit the most suitable vehicles to collect sensing data according to their daily movement patterns. In the second phase, a sensing task density-based participatory vehicle recruitment algorithm (STD-PVR) is proposed to reduce the computational complexity of matching vehicles with uncompleted sensing tasks. It is designed to recruit vehicles to arrive at designated locations to complete the sensing tasks within a given budget. Extensive evaluations based on a real-world dataset show that HR-DLVCS achieves higher sensing task completion rate than other baseline approaches in a variety of settings.