<p>Multi-object tracking requires accurately identifying and tracking multiple targets over long periods. However, tracking performance is highly susceptible to various factors, such as target deformation, occlusion, etc. Meanwhile, most existing MOT models perform simple aggregation and classification of target features, ignoring the inherent differences and connections between detection and re-identification. This often leads to frequent identity switches. To address the above issues, we propose our tracker IFMOT, a simple and efficient network that combines an interactive perception network with feature optimization. Specifically, we propose an interactive perception network with a multi-head cross-attention mechanism design to alleviate feature conflicts. And then, we introduce a feature optimization module that refines the target representation to improve the extraction capability of feature embeddings. Furthermore, a feature integration similarity matrix is used to comprehensively assess the similarity between objects and handle unreliable similarity matching. Experiments on the MOT16, MOT17, MOT20 and Dancetrack datasets show that the proposed method achieves a higher accuracy while keeping the tracking speed, in contrast to other state-of-the-art trackers.</p>