With the rapid increase in urban parking demand, smart parking technologies have emerged as a key solution for improving parking experience. Traditional approaches that rely on fixed sensors to monitor parking availability often incur high deployment and maintenance costs. Mobile Crowdsensing (MCS) offers a cost-effective alternative by utilizing taxis to dynamically collect parking space information as they traverse city streets. However, intelligently selecting which taxis to equip with sensing devices is critical for ensuring high-quality data. The inherent uncertainty in taxi trajectories poses significant challenges to traditional participant selection strategies, leading to performance degradation and instability in real-world deployments. To address this issue, we propose E2RL, a taxi selection method based on embedding representation learning and reinforcement learning. E2RL captures the behavioral patterns of taxis from historical trajectory data and learns an optimal taxi selection strategy through reinforcement learning. Extensive experiments conducted on the San Francisco taxi trajectory dataset and a real-world parking availability dataset demonstrate that our method significantly outperforms several baselines in terms of sensing quality and robustness.