Yolo Based Real-Time Object Detection for Vehicles

Accurate and quick object identification is a crucial part of autonomous driving since autonomous cars can operate safely and detect objects on the road. Object identification is a complex process that can generate false positives and negatives, impacting road safety. To that end, this study suggested an enhanced Yolov5-based quick and accurate object detection for self-driving cars. Firstly, training-inference decoupling with structural re-computerization (Rep) improved the YOLOv5 method by increasing the model's speed and accuracy. Furthermore, the neural architecture search approach was incorporated into the multi-branch re-computerization module to eliminate unnecessary branches during the training phase, improving training accuracy and efficiency. The coordinate attention method was integrated with each detection layer, while a small object detection layer was added at the end of the layer to increase the layer's identification rate.