This　paper　proposes　a　multi-target　tracking　and　trajectory　prediction　algorithm　for　dynamic　scenes.　By　combining　deep　learning　with　generative　adversarial　networks　(GAN),　efficient　target　detection　and　motion　prediction　in　complex　environments　are　achieved.　The　algorithm　first　uses　convolutional　neural　networks　and　long　short　-term　memory　networks　to　extract　the　spatiotemporal　features　of　dynamic　scenes,　then　uses　the　YOLO　detector　combined　with　Kalman　filtering　and　data　association　technology　to　generate　target　trajectories,　and　finally　predicts　future　trajectories　through　a　GAN-based　model.　Experimental　results　show　that　this　method　performs　well　in　multi-target　tracking　accuracy,　identity　retention　rate,　and　trajectory　prediction　error,　especially　in　target　occlusion　and　high-density　interaction　scenes.　It　shows　strong　robustness.　The　algorithm　effectively　overcomes　the　shortcomings　of　traditional　methods　in　nonlinear　motion　and　complex　dynamic　environments,　and　provides　high-performance　technical　support　for　fields　such　as　intelligent　monitoring　and　autonomous　driving.　©　2025　SPIE.