TinyThrow - Improved Lightweight Real-Time High-Rise Littering Object Detection Algorithm

Due to the high resolution and small object target in high-rise littering scene, the current object detection algorithms have problems of poor real-time detection and low detection accuracy. An improved lightweight real-time high-rise littering detection algorithm based on YOLOv5, TinyThrow, has been proposed to overcome the above challenges. The contribution of this paper is twofold: 1) A self-attention mechanism fuses traditional convolution module C3Trans is included to strengthen the functionality for global feature filter of our method. 2) An improved fast spatial channel attention mechanism module FCBAM is adopted to improve our algorithm’s ability of locating and identifying key targets. The results of experiments show that TinyThrow algorithm performs detection at a rate of 37.3 frames per second and mAP@.5 of 85.5% on an only small weight file of 3.9 MB, which is 4.5% higher than the original algorithm, meeting the requirements of lightweight real-time high-rise littering object detection task.