Modern　multi-object　tracking　methods　tend　to　focus　on　similarity　metrics　for　instances.　Existing　methods　adopt　the　mainstream　triplet　loss　to　help　deep　networks　learn　instance　similarity.　However,　triplet　loss　easily　ignores　samples　with　abnormal　similarity　and　difficult　samples;　thus,　the　networks　have　limited　ability　to　extract　object　features　for　similarity　calculation.　In　this　paper,　we　propose　a　novel　focal　triplet　loss　(FTL)　to　solve　such　problems.　FTL　pays　considerable　attention　to　objects　of　the　same　ID　with　low　similarity　and　different　IDs　with　high　similarity.　In　addition,　FTL　can　assign　additional　reasonable　loss　weights　to　difficult　samples,　making　the　networks　extract　more　discriminative　feature　embeddings　to　calculate　instance　similarity　and　improve　tracking　accuracy.　We　also　propose　a　new　joint　detection　and　embedding　model　integrating　a　detector　and　an　embedding　model　into　a　single　network.　Results　show　that　our　method　achieves　state-of-the-art　performance,　specifically　MOTA　of　87.28%　on　the　KITTI　dataset.　©　2021　IEEE.