The　key　to　person　re-identification　(Re-ID)　is　how　to　extract　a　representative　and　robust　depth　feature　of　the　person,　which　requires　the　model　to　pay　attention　to　both　global　contour　information　and　local　detailed　features.　To　extract　more　representative　features,　an　effective　method　is　to　build　a　multi-branch　deep　model　by　duplicating　the　backbone　structure.　This　method　usually　severely　increases　the　computational　cost,　and　continuous　convolution　and　pooling　operations　cause　the　loss　of　detailed　information.　This　paper　proposes　a　lightweight　multi-scale　feature　pyramid　structure,　which　extracts　features　from　network　layers　of　different　scales　and　aggregates　them　to　supplement　spatial　detail　information.　Meanwhile,　this　paper　adopts　a　pair　of　complementary　attention　modules,　which　pay　attention　to　the　discriminative　areas　of　person　features　by　focusing　on　channel　aggregation　and　position　perception,　respectively.　In　addition,　this　paper　proposes　a　multi-level　orthogonal　regularization　method　to　further　enhance　the　diversity　of　features.　The　experimental　results　show　that　the　mAP　of　this　method　on　the　Market1501　dataset　reaches　91.6%.　The　proposed　method　outperforms　state-of-the-art　methods　and　along　with　lower　complexity.