AIlib2/AI.py

import cv2,os,time,json
from models.experimental import attempt_load
from segutils.segmodel import SegModel,get_largest_contours
from segutils.trtUtils import segtrtEval,yolov5Trtforward,OcrTrtForward
from segutils.trafficUtils import tracfficAccidentMixFunction


from utils.torch_utils import select_device
from utilsK.queRiver import get_labelnames,get_label_arrays,post_process_,img_pad,draw_painting_joint,detectDraw,getDetections,getDetectionsFromPreds
from trackUtils.sort import moving_average_wang

from utils.datasets import letterbox
import numpy as np
import torch
import math
from PIL import Image
import torch.nn.functional as F
from copy import deepcopy
from scipy import interpolate

def xywh2xyxy(box,iW=None,iH=None):
    xc,yc,w,h = box[0:4]
    x0 =max(0, xc-w/2.0)
    x1 =min(1, xc+w/2.0)
    y0=max(0, yc-h/2.0)
    y1=min(1,yc+h/2.0)
    if iW: x0,x1 = x0*iW,x1*iW
    if iH: y0,y1 = y0*iH,y1*iH
    return [x0,y0,x1,y1]    
    

def get_ms(t2,t1):
    return (t2-t1)*1000.0
def get_postProcess_para(parfile):
    with open(parfile) as fp:
        par = json.load(fp)
    assert 'post_process' in par.keys(), ' parfile has not key word:post_process'   
    parPost=par['post_process']  
    
    return parPost["conf_thres"],parPost["iou_thres"],parPost["classes"],parPost["rainbows"]  
def get_postProcess_para_dic(parfile): 
    with open(parfile) as fp:
        par = json.load(fp)
    parPost=par['post_process']  
    return parPost
def AI_process(im0s,model,segmodel,names,label_arraylist,rainbows,objectPar={ 'half':True,'device':'cuda:0' ,'conf_thres':0.25,'iou_thres':0.45,'allowedList':[0,1,2,3],'segRegionCnt':1, 'trtFlag_det':False,'trtFlag_seg':False }, font={ 'line_thickness':None, 'fontSize':None,'boxLine_thickness':None,'waterLineColor':(0,255,255),'waterLineWidth':3} ,segPar={'modelSize':(640,360),'mean':(0.485, 0.456, 0.406),'std' :(0.229, 0.224, 0.225),'numpy':False, 'RGB_convert_first':True},mode='others',postPar=None):

    #输入参数
    #   im0s---原始图像列表
    #   model---检测模型，segmodel---分割模型（如若没有用到，则为None）
    #
    #输出：两个元素（列表，字符）构成的元组，[im0s[0],im0,det_xywh,iframe],strout
    #   [im0s[0],im0,det_xywh,iframe]中，
    #   im0s[0]--原始图像，im0--AI处理后的图像,iframe--帧号/暂时不需用到。
    #   det_xywh--检测结果，是一个列表。
    #       其中每一个元素表示一个目标构成如：[ xc,yc,w,h, float(conf_c),float(cls_c) ] ,2023.08.03修改输出格式
    #       #cls_c--类别，如0,1,2,3;  xc,yc,w,h--中心点坐标及宽；conf_c--得分, 取值范围在0-1之间
    #   #strout---统计AI处理个环节的时间
    # Letterbox

    half,device,conf_thres,iou_thres,allowedList = objectPar['half'],objectPar['device'],objectPar['conf_thres'],objectPar['iou_thres'],objectPar['allowedList']
    trtFlag_det,trtFlag_seg,segRegionCnt =   objectPar['trtFlag_det'],objectPar['trtFlag_seg'],objectPar['segRegionCnt']
    if 'ovlap_thres_crossCategory' in objectPar.keys():
        ovlap_thres = objectPar['ovlap_thres_crossCategory']
    else:
        ovlap_thres = None
    
    time0=time.time() 
    if trtFlag_det:
        img, padInfos = img_pad(im0s[0], size=(640,640,3))  ;img = [img]
    else:
        #print('####line72:',im0s[0][10:12,10:12,2])
        img = [letterbox(x, 640, auto=True, stride=32)[0] for x in im0s];padInfos=None
        #print('####line74:',img[0][10:12,10:12,2])
    # Stack
    img = np.stack(img, 0)
    # Convert
    img = img[:, :, :, ::-1].transpose(0, 3, 1, 2)  # BGR to RGB, to bsx3x416x416
    img = np.ascontiguousarray(img)  
    
    
    img = torch.from_numpy(img).to(device)
    img = img.half() if half else img.float()  # uint8 to fp16/32
    img /= 255.0
    time01=time.time()


    if segmodel: 
        seg_pred,segstr = segmodel.eval(im0s[0] )
        segFlag=True
    else:
        seg_pred = None;segFlag=False;segstr='Not implemented'

    time1=time.time()  
    if   trtFlag_det:  
        pred = yolov5Trtforward(model,img)
    else:
        #print('####line96:',img[0,0,10:12,10:12])
        pred = model(img,augment=False)[0]
       
    time2=time.time()


    p_result, timeOut = getDetectionsFromPreds(pred,img,im0s[0],conf_thres=conf_thres,iou_thres=iou_thres,ovlap_thres=ovlap_thres,padInfos=padInfos) 

    #if  mode=='highWay3.0':
    #if segmodel:
    if segPar and  segPar['mixFunction']['function']:

        mixFunction = segPar['mixFunction']['function'];H,W = im0s[0].shape[0:2]
        parMix = segPar['mixFunction']['pars'];#print('###line117:',parMix,p_result[2])
        parMix['imgSize'] = (W,H)
        #print(' -----------line110:  ',p_result[2] ,'\n', seg_pred)
        p_result[2] , timeMixPost= mixFunction(p_result[2], seg_pred, pars=parMix )
        #print(' -----------line112:  ',p_result[2] )
        p_result.append(seg_pred)

    else:
        timeMixPost=':0 ms'
    #print('#### line121: segstr:%s    timeMixPost:%s    timeOut:%s'%( segstr.strip(), timeMixPost,timeOut  ))
    time_info = 'letterbox:%.1f, seg:%.1f , infer:%.1f,%s, seginfo:%s  ,timeMixPost:%s '%( (time01-time0)*1000, (time1-time01)*1000 ,(time2-time1)*1000,timeOut , segstr.strip(),timeMixPost )
    #if mode=='highWay3.0':
        

    return p_result,time_info
       
def AI_process_forest(im0s,model,segmodel,names,label_arraylist,rainbows,half=True,device=' cuda:0',conf_thres=0.25, iou_thres=0.45,allowedList=[0,1,2,3], font={ 'line_thickness':None, 'fontSize':None,'boxLine_thickness':None,'waterLineColor':(0,255,255),'waterLineWidth':3} ,trtFlag_det=False,SecNms=None):
    #输入参数
    #   im0s---原始图像列表
    #   model---检测模型，segmodel---分割模型（如若没有用到，则为None）
    #输出：两个元素（列表，字符）构成的元组，[im0s[0],im0,det_xywh,iframe],strout
    #   [im0s[0],im0,det_xywh,iframe]中，
    #   im0s[0]--原始图像，im0--AI处理后的图像,iframe--帧号/暂时不需用到。
    #   det_xywh--检测结果，是一个列表。
    #       其中每一个元素表示一个目标构成如：[ xc,yc,w,h, float(conf_c),float(cls_c)],#2023.08.03,修改输出格式
    #       #cls_c--类别，如0,1,2,3;  xc,yc,w,h--中心点坐标及宽；conf_c--得分, 取值范围在0-1之间
    #   #strout---统计AI处理个环节的时间
    
     # Letterbox
    time0=time.time() 
    if trtFlag_det:
        img, padInfos = img_pad(im0s[0], size=(640,640,3))  ;img = [img]
    else:
        img = [letterbox(x, 640, auto=True, stride=32)[0] for x in im0s];padInfos=None
    #img = [letterbox(x, 640, auto=True, stride=32)[0] for x in im0s]
    # Stack
    img = np.stack(img, 0)
    # Convert
    img = img[:, :, :, ::-1].transpose(0, 3, 1, 2)  # BGR to RGB, to bsx3x416x416
    img = np.ascontiguousarray(img)  

    img = torch.from_numpy(img).to(device)
    img = img.half() if half else img.float()  # uint8 to fp16/32

    img /= 255.0  # 0 - 255 to 0.0 - 1.0
    if segmodel:
        seg_pred,segstr = segmodel.eval(im0s[0] )
        segFlag=True
    else:
        seg_pred = None;segFlag=False
    time1=time.time()     
    pred = yolov5Trtforward(model,img) if trtFlag_det else model(img,augment=False)[0]
    
        
    time2=time.time()
    datas = [[''], img, im0s, None,pred,seg_pred,10]        

    ObjectPar={ 'object_config':allowedList, 'slopeIndex':[] ,'segmodel':segFlag,'segRegionCnt':0 }
    p_result,timeOut = post_process_(datas,conf_thres, iou_thres,names,label_arraylist,rainbows,10,ObjectPar=ObjectPar,font=font,padInfos=padInfos,ovlap_thres=SecNms)
    #print('###line274:',p_result[2])
    #p_result,timeOut = post_process_(datas,conf_thres, iou_thres,names,label_arraylist,rainbows,10,object_config=allowedList,segmodel=segFlag,font=font,padInfos=padInfos)
    time_info = 'letterbox:%.1f, infer:%.1f, '%( (time1-time0)*1000,(time2-time1)*1000 )
    return p_result,time_info+timeOut
    

def AI_det_track( im0s_in,modelPar,processPar,sort_tracker,segPar=None):
    im0s,iframe=im0s_in[0],im0s_in[1]
    model =  modelPar['det_Model']  
    segmodel =   modelPar['seg_Model']    
    half,device,conf_thres, iou_thres,trtFlag_det = processPar['half'], processPar['device'], processPar['conf_thres'], processPar['iou_thres'],processPar['trtFlag_det']
    iou2nd = processPar['iou2nd']
    time0=time.time() 

    if trtFlag_det:
        img, padInfos = img_pad(im0s[0], size=(640,640,3))  ;img = [img]
    else:
        img = [letterbox(x, 640, auto=True, stride=32)[0] for x in im0s];padInfos=None
    img = np.stack(img, 0)
    # Convert
    img = img[:, :, :, ::-1].transpose(0, 3, 1, 2)  # BGR to RGB, to bsx3x416x416
    img = np.ascontiguousarray(img)  

    img = torch.from_numpy(img).to(device)
    img = img.half() if half else img.float()  # uint8 to fp16/32

    img /= 255.0  # 0 - 255 to 0.0 - 1.0
    
    seg_pred = None;segFlag=False
    time1=time.time()     
    pred = yolov5Trtforward(model,img) if trtFlag_det else model(img,augment=False)[0]
       
    time2=time.time()

    #p_result,timeOut = getDetections(datas,conf_thres, iou_thres,names,label_arraylist,rainbows,10,ObjectPar=ObjectPar,font=font,padInfos=padInfos)
    p_result, timeOut = getDetectionsFromPreds(pred,img,im0s[0],conf_thres=conf_thres,iou_thres=iou_thres,ovlap_thres=iou2nd,padInfos=padInfos) 
    
    if segmodel: 
        seg_pred,segstr = segmodel.eval(im0s[0] )
        segFlag=True
    else:
        seg_pred = None;segFlag=False;segstr='No segmodel'
    

    if segPar and segPar['mixFunction']['function']:
        mixFunction = segPar['mixFunction']['function']
        
        H,W = im0s[0].shape[0:2]
        parMix = segPar['mixFunction']['pars'];#print('###line117:',parMix,p_result[2])
        parMix['imgSize'] = (W,H)
        
        
        p_result[2],timeInfos_post = mixFunction(p_result[2], seg_pred, pars=parMix )
        timeInfos_seg_post = 'segInfer:%s ,postMixProcess:%s'%(  segstr, timeInfos_post  )
    else:
        timeInfos_seg_post = ' '
    '''    
    if segmodel:
        timeS1=time.time()
        #seg_pred,segstr = segtrtEval(segmodel,im0s[0],par=segPar) if segPar['trtFlag_seg'] else segmodel.eval(im0s[0] )
        seg_pred,segstr = segmodel.eval(im0s[0] )
        timeS2=time.time()
        mixFunction = segPar['mixFunction']['function']
     
        p_result[2],timeInfos_post = mixFunction(p_result[2], seg_pred, pars=segPar['mixFunction']['pars']  )

        timeInfos_seg_post = 'segInfer:%.1f ,postProcess:%s'%(  (timeS2-timeS1)*1000, timeInfos_post  )
        
    else:
        timeInfos_seg_post = ' '
        #print('######line341:',seg_pred.shape,np.max(seg_pred),np.min(seg_pred) , len(p_result[2]) )  
    '''        
    time_info = 'letterbox:%.1f, detinfer:%.1f, '%( (time1-time0)*1000,(time2-time1)*1000 )
    
    if sort_tracker:
        #在这里增加设置调用追踪器的频率
        #..................USE TRACK FUNCTION....................
        #pass an empty array to sort
        dets_to_sort = np.empty((0,7), dtype=np.float32)
        
        # NOTE: We send in detected object class too
        #for detclass,x1,y1,x2,y2,conf in p_result[2]:
        for x1,y1,x2,y2,conf, detclass in p_result[2]:
            #print('#######line342:',x1,y1,x2,y2,img.shape,[x1, y1, x2, y2, conf, detclass,iframe])
            dets_to_sort = np.vstack((dets_to_sort, 
                        np.array([x1, y1, x2, y2, conf, detclass,iframe],dtype=np.float32)   ))
        
        # Run SORT
        tracked_dets = deepcopy(sort_tracker.update(dets_to_sort) ) 
        tracks =sort_tracker.getTrackers()  
        p_result.append(tracked_dets)  ###index=4          
        p_result.append(tracks)  ###index=5  
    
    return p_result,time_info+timeOut+timeInfos_seg_post
def AI_det_track_batch(imgarray_list, iframe_list ,modelPar,processPar,sort_tracker,trackPar,segPar=None):  
    '''
    输入：
        imgarray_list--图像列表
        iframe_list -- 帧号列表
        modelPar--模型参数,字典,modelPar={'det_Model':,'seg_Model':}
        processPar--字典，存放检测相关参数，'half', 'device', 'conf_thres', 'iou_thres','trtFlag_det'
        sort_tracker--对象，初始化的跟踪对象。为了保持一致，即使是单帧也要有。
        trackPar--跟踪参数，关键字包括：det_cnt，windowsize
        segPar--None,分割模型相关参数。如果用不到，则为None
    输入：retResults,timeInfos
        retResults：list
            retResults[0]--imgarray_list
            retResults[1]--所有结果用numpy格式，所有的检测结果，包括8类，每列分别是x1, y1, x2, y2, conf, detclass,iframe,trackId
            retResults[2]--所有结果用list表示,其中每一个元素为一个list，表示每一帧的检测结果，每一个结果是由多个list构成，每个list表示一个框，格式为[ x0 ,y0  ,x1  ,y1  ,conf, cls ,ifrmae,trackId ]，如 retResults[2][j][k]表示第j帧的第k个框。2023.08.03,修改输出格式
    '''
     
    det_cnt,windowsize = trackPar['det_cnt']  ,trackPar['windowsize']
    trackers_dic={}
    index_list = list(range( 0, len(iframe_list) ,det_cnt  ));
    if len(index_list)>1 and  index_list[-1]!= iframe_list[-1]:
        index_list.append( len(iframe_list) - 1  )
        
    if len(imgarray_list)==1:   #如果是单帧图片，则不用跟踪
        retResults = []
        p_result,timeOut =  AI_det_track( [ [imgarray_list[0]]  ,iframe_list[0] ],modelPar,processPar,None,segPar )
        ##下面4行内容只是为了保持格式一致
        detArray = np.array(p_result[2])
        #print('##line371:',detArray)
        if len(p_result[2])==0:res=[]
        else:    
            cnt = detArray.shape[0];trackIds=np.zeros((cnt,1));iframes = np.zeros((cnt,1)) + iframe_list[0]
           
            #detArray = np.hstack( (detArray[:,1:5],  detArray[:,5:6] ,detArray[:,0:1],iframes, trackIds )  )
            detArray = np.hstack( (detArray[:,0:4],  detArray[:,4:6] ,iframes, trackIds )  ) ##2023.08.03 修改输入格式
            res = [[ b[0],b[1],b[2],b[3],b[4],b[5],b[6],b[7] ] for b in detArray ]  
        retResults=[imgarray_list,detArray,res ]    
        #print('##line380:',retResults[2])
        return retResults,timeOut
    
    else:
        t0 = time.time()
        timeInfos_track=''
        for iframe_index, index_frame in enumerate(index_list):
            p_result,timeOut =  AI_det_track( [ [imgarray_list[index_frame]]  ,iframe_list[index_frame] ],modelPar,processPar,sort_tracker,segPar )
            timeInfos_track='%s:%s'%(timeInfos_track,timeOut)
            
            for tracker in  p_result[5]:
                trackers_dic[tracker.id]=deepcopy(tracker)     
        t1 = time.time()
        
        track_det_result = np.empty((0,8))
        for trackId in trackers_dic.keys():
            tracker  = trackers_dic[trackId]
            bbox_history = np.array(tracker.bbox_history)
            if len(bbox_history)<2: continue
            ###把(x0,y0,x1,y1)转换成(xc,yc,w,h)
            xcs_ycs = (bbox_history[:,0:2] + bbox_history[:,2:4] )/2
            whs = bbox_history[:,2:4] - bbox_history[:,0:2] 
            bbox_history[:,0:2] = xcs_ycs;bbox_history[:,2:4] = whs;
            
            arrays_box = bbox_history[:,0:7].transpose();frames=bbox_history[:,6]
            #frame_min--表示该批次图片的起始帧，如该批次是[1,100],则frame_min=1，[101,200]--frame_min=101
            #frames[0]--表示该目标出现的起始帧，如[1,11,21,31,41],则frames[0]=1，frames[0]可能会在frame_min之前出现，即一个横跨了多个批次。
     
            ##如果要最好化插值范围，则取内区间[frame_min,则frame_max ]和[frames[0],frames[-1] ]的交集
            #inter_frame_min = int(max(frame_min, frames[0])); inter_frame_max = int(min( frame_max, frames[-1] )) ##
            
            ##如果要求得到完整的目标轨迹，则插值区间要以目标出现的起始点为准
            inter_frame_min=int(frames[0]);inter_frame_max=int(frames[-1])
            new_frames= np.linspace(inter_frame_min,inter_frame_max,inter_frame_max-inter_frame_min+1 )
            f_linear = interpolate.interp1d(frames,arrays_box); interpolation_x0s = (f_linear(new_frames)).transpose()
            move_cnt_use =(len(interpolation_x0s)+1)//2*2-1  if len(interpolation_x0s)<windowsize else windowsize
            for im in range(4):
                interpolation_x0s[:,im] = moving_average_wang(interpolation_x0s[:,im],move_cnt_use )
            
            cnt = inter_frame_max-inter_frame_min+1; trackIds = np.zeros((cnt,1)) + trackId
            interpolation_x0s = np.hstack( (interpolation_x0s, trackIds )  )
            track_det_result =  np.vstack(( track_det_result, interpolation_x0s)  )
            #print('#####line116:',trackId,frame_min,frame_max,'----------',interpolation_x0s.shape,track_det_result.shape ,'-----')
            
        ##将[xc,yc,w,h]转为[x0,y0,x1,y1] 
        x0s =   track_det_result[:,0] -  track_det_result[:,2]/2 ; x1s =   track_det_result[:,0] +  track_det_result[:,2]/2     
        y0s =   track_det_result[:,1] -  track_det_result[:,3]/2 ; y1s =   track_det_result[:,1] +  track_det_result[:,3]/2 
        track_det_result[:,0] = x0s; track_det_result[:,1] = y0s; 
        track_det_result[:,2] = x1s; track_det_result[:,3] = y1s;
        detResults=[]
        for iiframe in iframe_list:
            boxes_oneFrame = track_det_result[ track_det_result[:,6]==iiframe   ]
            res = [[ b[0],b[1],b[2],b[3],b[4],b[5],b[6],b[7] ] for b in boxes_oneFrame ]
            #[ x0 ,y0  ,x1  ,y1  ,conf,cls,ifrmae,trackId ]
            #[ifrmae,  x0 ,y0  ,x1  ,y1  ,conf,cls,trackId ]
            detResults.append( res )
           
  
        retResults=[imgarray_list,track_det_result,detResults ]    
        t2 = time.time()
        timeInfos = 'detTrack:%.1f TrackPost:%.1f, %s'%(get_ms(t1,t0),get_ms(t2,t1), timeInfos_track )
        return retResults,timeInfos
                
    
def ocr_process(pars):

    img_patch,engine,context,converter,AlignCollate_normal,device=pars[0:6]
    time1 = time.time()
    img_tensor = AlignCollate_normal([ Image.fromarray(img_patch,'L') ])
    img_input = img_tensor.to('cuda:0')
    time2 = time.time()
    
    preds,trtstr=OcrTrtForward(engine,[img_input],context)
    time3 = time.time()
    
    batch_size = preds.size(0)
    preds_size = torch.IntTensor([preds.size(1)] * batch_size)

    ######## filter ignore_char, rebalance
    preds_prob = F.softmax(preds, dim=2)
    preds_prob = preds_prob.cpu().detach().numpy()
    pred_norm = preds_prob.sum(axis=2)
    preds_prob = preds_prob/np.expand_dims(pred_norm, axis=-1)
    preds_prob = torch.from_numpy(preds_prob).float().to(device)
    _, preds_index = preds_prob.max(2)
    preds_index = preds_index.view(-1)  
    time4 = time.time()
    preds_str = converter.decode_greedy(preds_index.data.cpu().detach().numpy(), preds_size.data)
    time5 = time.time()
    
    info_str= ('pre-process:%.2f  TRTforward:%.2f (%s)  postProcess:%2.f  decoder:%.2f, Total:%.2f , pred:%s'%(get_ms(time2,time1 ),get_ms(time3,time2 ),trtstr, get_ms(time4,time3 ), get_ms(time5,time4 ),  get_ms(time5,time1 ),  preds_str  )  )
    return  preds_str,info_str
def main():
    ##预先设置的参数
    device_='1'  ##选定模型，可选 cpu,'0','1'
    
    ##以下参数目前不可改
    Detweights = "weights/yolov5/class5/best_5classes.pt"
    seg_nclass = 2 
    Segweights = "weights/BiSeNet/checkpoint.pth"
    conf_thres,iou_thres,classes= 0.25,0.45,5
    labelnames = "weights/yolov5/class5/labelnames.json"
    rainbows = [ [0,0,255],[0,255,0],[255,0,0],[255,0,255],[255,255,0],[255,129,0],[255,0,127],[127,255,0],[0,255,127],[0,127,255],[127,0,255],[255,127,255],[255,255,127],[127,255,255],[0,255,255],[255,127,255],[127,255,255],  [0,127,0],[0,0,127],[0,255,255]]
    allowedList=[0,1,2,3]
    
    
    ##加载模型，准备好显示字符
    device = select_device(device_)
    names=get_labelnames(labelnames)
    label_arraylist = get_label_arrays(names,rainbows,outfontsize=40,fontpath="conf/platech.ttf")
    half = device.type != 'cpu'  # half precision only supported on CUDA
    model = attempt_load(Detweights, map_location=device)  # load FP32 model
    if half: model.half()
    segmodel = SegModel(nclass=seg_nclass,weights=Segweights,device=device)
    

    ##图像测试
    #url='images/examples/20220624_响水河_12300_1621.jpg'
    impth = 'images/examples/'
    outpth = 'images/results/'
    folders = os.listdir(impth)
    for i in range(len(folders)):
        imgpath = os.path.join(impth, folders[i])
        im0s=[cv2.imread(imgpath)]
        time00 = time.time()
        p_result,timeOut = AI_process(im0s,model,segmodel,names,label_arraylist,rainbows,half,device,conf_thres, iou_thres,allowedList,fontSize=1.0)
        time11 = time.time()
        image_array = p_result[1]
        cv2.imwrite(  os.path.join( outpth,folders[i] ) ,image_array )
        #print('----process:%s'%(folders[i]), (time.time() - time11) * 1000)

    
    
    
        
if __name__=="__main__":
    main()