Objective　The　movement　of　sediment　particles　significantly　affects　riverbed　evolution,　establishing　it　as　a　central　concern　and　ongoing　challenge　in　fluvial　dynamics　research.　Although　image　processing　methods　provide　efficient　means　for　acquiring　and　analyzing　data　on　sediment　transport　characteristics,　their　accuracy　is　often　compromised　by　water　waves,　bubbles,　and　threshold　errors.　Therefore,　continuous　improvements　and　refinements　remain　essential　to　ensure　the　acquisition　of　accurate　and　reliable　particle　state　data.　This　study　integrates　deep　learning　networks　with　existing　image　processing　techniques　to　enable　more　precise　and　comprehensive　identification　of　suspended　sediment　particles.　It　further　investigates　the　relationship　between　turbulent　coherent　structures　and　the　intensity　of　particle　movement,　clarifying　the　mechanism　through　which　turbulent　coherent　structures　influence　sediment　transport.　Methods　The　optimization　algorithm　developed　in　this　study　aims　to　maximize　the　detection　of　moving　particles,　providing　more　accurate　data　to　support　understanding　sediment　transport　patterns　at　the　particle　scale　and　their　association　with　turbulent　coherent　structures.　This　research　provides　new　insights　for　advanced　measurement　techniques　and　the　exploration　of　sediment　transport　mechanisms.　Bedload　equilibrium　sediment　transport　experiments　are　conducted　under　medium　to　low　flow　conditions　(Θ　=　0.052　to　0.071).　High-speed　cameras　are　utilized　to　capture　images　of　bedload　particles　during　water　flow　scouring　processes.　An　optimized　method　for　identifying　bedload　particle　motion　is　proposed　by　combining　the　grayscale　subtraction　method　with　deep　learning　techniques.　The　grayscale　subtraction　method　identifies　regions　of　particle　motion　by　calculating　differences　in　grayscale　values　between　consecutive　frames　and　separately　analyzing　the　centroids　of　moving　particles　in　each　frame.　However,　because　this　method　depends　solely　on　grayscale　variations,　it　presents　limitations　in　identifying　regions　with　minor　grayscale　changes.　The　YOLOv5　(you　only　look　once)　method　is　designed　to　rapidly　and　accurately　detect　specific　target　objects　and　their　locations　in　images　after　training　on　a　sampled　dataset.　The　YOLOv5　algorithm　adopted　in　this　study　excels　at　detecting　small　targets　and　provides　multi-scale　detection,　strong　versatility,　fast　training,　inference　speeds,　and　adaptable　fine-tuning　capabilities.　The　YOLOv5　deep　learning　network　structure　is　enhanced　by　improving　convolutional　blocks,　incorporating　attention　mechanisms,　and　optimizing　loss　function　processing,　boosting　the　detector＇s　overall　performance　in　accurately　capturing　the　motion　of　particles　over　short　distances.　The　particle　tracking　velocimetry　method　and　Kalman　filtering　algorithm　are　employed　to　calculate　the　trajectories　of　bedload　particles.　Results　and　Discussion　The　improved　YOLOv5　model　demonstrates　significant　enhancements　in　loss　function　handling,　detection　accuracy,　and　precision.　The　detection　accuracy　of　the　improved　model　for　suspended　sediment　particles　reaches　94.9%,　with　a　2.3%　increase　in　average　precision　and　respective　gains　of　1.1%　and　1.0%　in　precision　and　recall　rates.　The　weighted　harmonic　mean　of　the　comprehensive　verification　index,　F1　score,　increases　by　two　percentage　points.　This　enhanced　performance　in　practical　detection　surpasses　that　of　the　original　YOLOv5　model.　The　number　of　observed　particle　chains　increases　following　optimization　by　integrating　the　improved　YOLOv5　model　with　the　grayscale　subtraction　technique　for　detecting　particle　motion.　Analyses　of　cumulative　centroid　counts　and　particle　chain　node　counts　reveal　an　ascending　trend　as　the　number　of　frames　increases.　The　cumulative　centroid　count　and　particle　chain　node　count　obtained　through　the　optimization　method　remain　stable　at　approximately　59%　and　80%,　respectively,　contrasting　with　the　growth　percentages　of　the　individual　methods.　It　is　proposed　that　the　formation　of　sediment　particle　motion　bands　is　associated　with　Q2/Q4　bursting　events　of　coherent　turbulent　structures　based　on　the　results　of　particle　motion.　During　Q4　events,　the　average　flow　velocity　exceeds　that　observed　during　Q2　events.　Under　identical　water　depth　conditions,　the　shear　force　in　the　Q4　region　surpasses　that　in　the　Q2　region,　resulting　in　a　higher　concentration　of　sediment　particles　in　the　corresponding　Q4　region.　The　bed　surface　structure　exhibits　convex　grooves　in　the　Q4　region　and　concave　grooves　in　the　Q2　region,　extending　across　the　entire　bed　surface　in　the　spanwise　direction　of　the　channel.　Characteristics　of　coherent　turbulent　structures　provide　a　more　comprehensive　explanation　for　the　mechanism　underlying　the　formation　of　sediment　transport　belts.　Conclusion　This　study　concludes　the　following:　1)　The　grayscale　subtraction　technique　effectively　identifies　particles　with　significant　motion　distances,　while　deep　learning　methods　excel　at　recognizing　particles　with　smaller　motion　distances.　Through　comparative　analysis,　data　evaluation,　and　experimental　observations,　it　becomes　evident　that　the　integrated　algorithm,　which　combines　both　approaches,　enhances　the　accuracy　of　bedload　particle　and　trajectory　identification　under　moderate　to　low　flow　conditions.　2)Under　conditions　of　moderate　to　low　flow　intensity,　the　motion　intensity　of　bedload　sediment　particles　is　influenced　by　coherent　turbulent　structures,　resulting　in　a　laterally　banded　structure.　As　flow　intensity　increases,　the　banded　structure　becomes　sparser　and　wider.　However,　further　intensification　of　the　flow　leads　to　vigorous　turbulent　mixing,　which　weakens　the　coherent　turbulent　structures　and　ultimately　causes　the　banded　structure　to　disappear.　3)Overall,　sediment　particle　motion　primarily　concentrates　in　the　central　region　of　the　channel,　with　reduced　motion　observed　near　the　sidewalls　due　to　lower　flow　velocities　within　the　boundary　layer.　This　observation　aligns　with　practical　scenarios　and　hydraulic　theory.　In　addition,　the　morphology　of　the　sediment-streaky　structure　typically　exhibits　a　wider　middle　section　and　narrower　sides,　indicating　that　the　formation　of　the　banded　structure　is　primarily　influenced　by　large-scale　coherent　turbulent　structures　rather　than　secondary　flow　structures.　This　study　introduces　deep　learning　into　conventional　bedload　particle　motion　recognition,　improving　the　accuracy　of　bedload　particle　identification　from　a　higher-resolution　perspective.　It　addresses　the　challenge　of　detecting　multiple　moving　sediment　particles　and　provides　a　useful　reference　for　research　in　fluvial　dynamics.　©　2025　Sichuan　University.　All　rights　reserved.