ISiamDW：Improved SiamRPN Target Tracking Algorithm Based on Inverted Residual Mobile Block

Abstract The SiamRPN target tracking algorithm has received widespread attention due to its good balance between speed and performance. Its concise network structure makes it faster, but the tracking effect is not satisfactory. Especially in complex backgrounds, SiamRPN often loses targets. The SiamRPN algorithm lacks the ability to recognize interference from similar objects when using deep semantic features for tracking. Therefore, we have introduced an inverted residual mobile block for this purpose. The inverted residual mobile block utilizes multi-head self-attention and depthwise separable convolution cascades to decompose and recombine the feature information extracted from the backbone Siamese network of SiamRPN, improving the model's generalization ability and avoiding the influence of similar targets in complex scenes which may result in incorrect tracking. In addition, we extensively use depthwise separable convolutions to reduce the number of parameters and improve tracking efficiency. The tracker achieved good results on three large tracking benchmarks, including OTB2015, UAV123, and VOT2016. Especially in UAV123, it has excellent performance and can be applied in unmanned aerial vehicle tracking and enemy search.