增加了单帧入侵判断及yoloV10

contrast/utils/calsimi.py

@@ -0,0 +1,216 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+"""
+Created on Mon Mar 31 16:25:43 2025
+
+@author: wqg
+"""
+import numpy as np
+from scipy.spatial.distance import cdist
+
+
+def get_topk_percent(data, k):
+    """
+    获取数据中最大的 k% 的元素
+    """
+    # 将数据转换为 NumPy 数组
+    if isinstance(data, list):
+        data = np.array(data)
+    
+    percentile = np.percentile(data, 100-k)
+    top_k_percent = data[data >= percentile]
+    
+    return top_k_percent
+def cluster(data, thresh=0.15):
+    # data = np.array([0.1, 0.13, 0.7, 0.2, 0.8, 0.52, 0.3, 0.7, 0.85, 0.58])
+    # data = np.array([0.1, 0.13, 0.2, 0.3])
+    # data = np.array([0.1])
+    
+    if isinstance(data, list):
+        data = np.array(data)
+    
+    data1 = np.sort(data)
+    cluter, Cluters,  = [data1[0]], []
+    for i in range(1, len(data1)):
+        if data1[i] - data1[i-1]< thresh:
+            cluter.append(data1[i])
+        else:
+            Cluters.append(cluter)
+            cluter = [data1[i]]
+    Cluters.append(cluter)
+    
+    clt_center = []
+    for clt in Cluters:
+        ## 是否应该在此处限制一个聚类中的最小轨迹样本数，应该将该因素放在轨迹分析中
+        # if len(clt)>=3:
+        #     clt_center.append(np.mean(clt))
+        clt_center.append(np.mean(clt))
+    
+    # print(clt_center)
+        
+    return clt_center
+
+def calsiml(feat1, feat2, topkp=75, cluth=0.15):
+    '''轨迹样本和标准特征集样本相似度的选择策略'''
+    matrix = 1 - cdist(feat1, feat2, 'cosine')
+    simi_max = []
+    for i in range(len(matrix)):
+        sim = np.mean(get_topk_percent(matrix[i, :], topkp))
+        simi_max.append(sim)
+    cltc_max = cluster(simi_max, cluth)
+    Simi = max(cltc_max) 
+    
+    ## cltc_max为空属于编程考虑不周，应予以排查解决 
+    # if len(cltc_max):
+    #     Simi = max(cltc_max) 
+    # else:
+    #     Simi = 0   #不应该走到该处
+
+    return Simi
+
+
+def calsimi_vs_stdfeat_new(event, stdfeat):
+    '''事件与标准库的对比策略
+       该比对策略是否可以拓展到事件与事件的比对？
+    '''
+    front_boxes = np.empty((0, 9), dtype=np.float64) ##和类doTracks兼容 
+    front_feats = np.empty((0, 256), dtype=np.float64) ##和类doTracks兼容 
+    for i in range(len(event.front_boxes)):
+        front_boxes = np.concatenate((front_boxes, event.front_boxes[i]), axis=0)
+        front_feats = np.concatenate((front_feats, event.front_feats[i]), axis=0)
+    
+    back_boxes = np.empty((0, 9), dtype=np.float64) ##和类doTracks兼容 
+    back_feats = np.empty((0, 256), dtype=np.float64) ##和类doTracks兼容 
+    for i in range(len(event.back_boxes)):
+        back_boxes = np.concatenate((back_boxes, event.back_boxes[i]), axis=0)
+        back_feats = np.concatenate((back_feats, event.back_feats[i]), axis=0)
+    
+    front_simi, back_simi = None, None
+    if len(front_feats):
+        front_simi = calsiml(front_feats, stdfeat)
+    if len(back_feats):
+        back_simi = calsiml(back_feats, stdfeat)
+
+    '''前后摄相似度融合策略'''
+    if len(front_feats) and len(back_feats):
+        diff_simi = abs(front_simi - back_simi) 
+        if diff_simi>0.15:
+            Similar = max([front_simi, back_simi])
+        else:
+            Similar = (front_simi+back_simi)/2
+    elif len(front_feats) and len(back_feats)==0:
+        Similar = front_simi
+    elif len(front_feats)==0 and len(back_feats):
+        Similar = back_simi
+    else:
+        Similar = None   # 在event.front_feats和event.back_feats同时为空时
+
+    return Similar, front_simi, back_simi
+
+
+
+
+
+def calsimi_vs_stdfeat(event, stdfeat):
+    evtfeat = event.feats_compose
+    if isinstance(event.feats_select, list):
+        if len(event.feats_select) and len(event.feats_select[0]):
+            evtfeat = event.feats_select[0]
+        else:
+            return None, None, None
+    else:
+        evtfeat = event.feats_select
+
+    if len(evtfeat)==0 or len(stdfeat)==0:
+        return None, None, None
+    
+    
+    evtfeat /= np.linalg.norm(evtfeat, axis=1)[:, None]
+    stdfeat /= np.linalg.norm(stdfeat, axis=1)[:, None]
+
+    matrix = 1 - cdist(evtfeat, stdfeat, 'cosine')
+    matrix[matrix < 0] = 0
+    
+    simi_mean = np.mean(matrix)
+    simi_max  = np.max(matrix)
+    stdfeatm = np.mean(stdfeat, axis=0, keepdims=True)
+    evtfeatm = np.mean(evtfeat, axis=0, keepdims=True)
+    simi_mfeat = 1- np.maximum(0.0, cdist(stdfeatm, evtfeatm, 'cosine'))
+    
+    return simi_mean, simi_max, simi_mfeat[0,0]
+
+
+def calsimi_vs_evts(evta, evtb, simType=1):
+    if simType==1:
+        if len(evta.feats_compose) and len(evtb.feats_compose):
+            feata = evta.feats_compose
+            featb = evtb.feats_compose
+            matrix = 1 - cdist(feata, featb, 'cosine')
+            similar = np.mean(matrix)
+        else:
+            similar = None
+        return similar
+    
+    if simType==2:
+        if len(evta.feats_compose) and len(evtb.feats_compose):
+            feata = evta.feats_compose
+            featb = evtb.feats_compose
+            matrix = 1 - cdist(feata, featb, 'cosine')
+            similar = np.max(matrix)
+        else:
+            similar = None
+        return similar
+
+    if simType==3:
+        if len(evta.feats_compose) and len(evtb.feats_compose):
+            feata = evta.feats_compose
+            featb = evtb.feats_compose
+            similar = calsiml(feata, featb)
+        else:
+            similar = None
+        return similar
+            
+    
+    ##1. the front feats of evta, evtb 
+    fr_feata = np.empty((0, 256), dtype=np.float64) ##和类doTracks兼容 
+    for i in range(len(evta.front_feats)):
+        fr_feata = np.concatenate((fr_feata, evta.front_feats[i]), axis=0)
+        
+    fr_featb = np.empty((0, 256), dtype=np.float64) ##和类doTracks兼容 
+    for i in range(len(evtb.front_feats)):
+        fr_featb = np.concatenate((fr_featb, evtb.front_feats[i]), axis=0)
+        
+    ##2. the back feats of evta, evtb     
+    bk_feata = np.empty((0, 256), dtype=np.float64) ##和类doTracks兼容 
+    for i in range(len(evta.back_feats)):
+        bk_feata = np.concatenate((bk_feata, evta.back_feats[i]), axis=0)
+   
+    bk_featb = np.empty((0, 256), dtype=np.float64) ##和类doTracks兼容 
+    for i in range(len(evtb.back_feats)):
+        bk_featb = np.concatenate((bk_featb, evtb.back_feats[i]), axis=0)
+    
+    
+    front_simi, back_simi = None, None
+    if len(fr_feata) and len(fr_featb):
+        front_simi = calsiml(fr_feata, fr_featb)
+    
+    if len(bk_feata) and len(bk_featb):
+        back_simi = calsiml(bk_feata, bk_featb)
+
+    '''前后摄相似度融合策略'''
+    if front_simi is not None and back_simi is not None:
+        diff_simi = abs(front_simi - back_simi) 
+        if diff_simi>0.15:
+            similar = max([front_simi, back_simi])
+        else:
+            similar = (front_simi+back_simi)/2
+    elif front_simi is not None and back_simi is None:
+        similar = front_simi
+    elif front_simi is None and back_simi is not None:
+        similar = back_simi
+    else:
+        similar = None   # 在event.front_feats和event.back_feats同时为空时
+
+    return similar
+
+