Object　tracking　serves　as　a　prerequisite　and　foundation　for　higher-level　driving　tasks　and　has　broad　application　prospects　in　various　fields,　including　intelligent　logistics　and　autonomous　driving.　In　this　paper,　we　propose　a　deep　reinforcement　learning-based　object　tracking　control　network　model,　that　incorporates　attention　and　Long　Short-Term　Memory　(LSTM)　mechanisms.　The　Asynchronous　Advantage　Actor-Critic　(A3C)　algorithm　is　employed　for　multi-threaded　synchronous　unsupervised　training　of　the　tracking　network　model,　resulting　in　end-to-end　dynamic　object　tracking　control.　Additionally,　the　Gradient-weighted　Class　Activation　Mapping　(Grad-CAM)　method　is　utilized　to　analyze　the　interpretability　of　the　network　model.　Experimental　results　demonstrate　that　by　introducing　the　attention　salience　mechanism　and　LSTM　temporal　mechanism,　the　network　can　effectively　focus　on　obstacle　and　target　locations,　thus　enhancing　its　attentional　capacity　and　interpretability.　The　trustworthiness　of　the　object　tracking　model　can　be　improved　in　terms　of　both　tracking　performance　and　interpretability.　©　2023　IEEE.