Applying Image Analysis to Build a Lightweight System for Blind Obstacles Detecting of Intelligent Wheelchairs

Intelligent wheelchair blind spot obstacle detection is an important issue for semi-enclosed special environments in elderly communities. However, the LiDAR- and 3D-point-cloud-based solutions are expensive, complex to deploy, and require significant computing resources and time. This paper proposed an improved YOLOV5 lightweight obstacle detection model, named GC-YOLO, and built an obstacle dataset that consists of incomplete target images captured in the blind spot view of the smart wheelchair. The feature extraction operations are simplified in the backbone and neck sections of GC-YOLO. The backbone network uses GhostConv in the GhostNet network to replace the ordinary convolution in the original feature extraction network, reducing the model size. Meanwhile, the CoordAttention is applied, aiming to reduce the loss of location information caused by GhostConv. Further, the neck stem section uses a combination module of the lighter SE Attention module and the GhostConv module to enhance the feature extraction capability. The experimental results show that the proposed GC-YOLO outperforms the YOLO5 in terms of model parameters, GFLOPS and F1. Compared with the YOLO5, the number of model parameters and GFLOPS are reduced by 38% and 49.7%, respectively. Additionally, the F1 of the proposed GC-YOLO is improved by 10% on the PASCAL VOC dataset. Moreover, the proposed GC-YOLO achieved mAP of 90% on the custom dataset.