An improved lightweight small object detection framework applied to real-time autonomous driving

Recent deep learning-based object detectors have shown compelling performance for the detection of large objects in autonomous driving applications. However, the detection of small objects like traffic signs and traffic lights is challenging owing to the complex nature of such objects. This article investigates how an existing object detector can be adjusted to address specific tasks and how these modifications can impact the detection of small objects. In particular, we explore and introduce architectural changes to the different components of the popular YOLOv5 model in order to improve its performance in the detection of small objects for autonomous driving. Initially, we propose group depthwise separable convolution as the improved convolution unit to replace standard convolution. We then integrate this unit to create the attention-based dilated CSP block. Lastly, this block is combined with several proposed modules, including the improved SPP, improved PANet, and improved information paths, to form our IS-YOLOv5 model. We also integrate kernel pruning on the network to accelerate the model deployment on vehicle-mounted mobile platform due to limited computing resources and real-time constraints. Specifically, we propose the versatile network pruning (VNP) technique based on Taylor criterion ranking to prune less-essential kernels in the network. We will show that our modifications barely increase the complexity but significantly improve the detection accuracy and speed. Compared to the conventional YOLOv5, the proposed IS-YOLOv5 model increases the mAP by 8.35% on the BDD100K dataset. Besides, our proposed model improves the detection speed in FPS by 3.10% compared to the YOLOv5 model. When using the VNP scheme, FPS is further increased by 52.14%, while the model size and complexity are reduced by 39.29% and 47.81%, with almost no change in mAP. Nevertheless, when compared to state-of-the-art models, IS-YOLOv5+VNP is found to be conducive to the deployment in autonomous driving systems.