PASSIM: Parking Slot Recognition Using Attentional Semantic Segmentation and Instance Matching

Autonomous Valet Parking (AVP) has become one of the most promising applications in autonomous driving, where parking slot recognition plays an essential role. However, most existing vision-based parking slot recognition methods are not fully applicable for AVP systems. To obtain accurate parking slot segmentation and location results simultaneously without much prior knowledge, this paper proposes a novel recognition method using an attentional semantic segmentation network and an instance matching algorithm, namely PASSIM. The proposed method first exploits Around View Monitor (AVM) images to obtain semantic features of parking slots, including entry lines, separation lines, and corner points. Two attention-based blocks are integrated with the decoding part of the semantic segmentation network to enhance spatial and channel-wise correlation. After that, semantic instances are generated by categorical clustering, and keypoints are extracted from them for feature matching and parking slot inference. The final results are evaluated on the public dataset PS2.0 and compared with current open-source parking slot recognition methods. Experiment shows that PASSIM achieves accurate segmentation with a mean intersection of union (mIoU) score of 0.807 and has a comparable performance with state-of-the-art detection methods by showing recall of 98.57% and precision of 99.03%. In addition, PASSIM outperforms other approaches in terms of location accuracy.