Auto parts UK的問題，透過圖書和論文來找解法和答案更準確安心。 我們找到下列各種有用的問答集和懶人包

基於目的地導向之道路潛在危險社交行為預測

為了解決Auto parts UK的問題，作者林明志 這樣論述：

本論文主要開發一套基於目的地導向之道路潛在危險社交行為預測，如行人或車輛無預期性的突然闖入車道、行人不遵守道路規則橫跨馬路等道路危險情境，藉由所發展的深度學習演算策略預測動態物件的短期軌跡，以進一步達到駕駛安全預警輔助系統之功效。首先，為了提取道路環境中動態物件一小段連續時間的辨識結果，故本論文主要是採用深度學習模型進行物件辨識，並於辨識後使用件追蹤演算法，以確保獲得的邊界框為同一行人、四輪車輛或者兩輪車輛。接著我們發展一套基於目的地導向之社交行為預測模型，並搭配自我迴歸訓練策略，以實現物件彼此之間的社交軌跡預測，其中該網路模型主要分成五大部分 (1)特徵提取器；(2)編碼器；(2)目的地導

向預測器；(3)條件變分自動編碼器；(4)解碼器。首先，透過特徵提取器由輸入資訊中提取動態物件與自車彼此間的距離、動態物件速度、動態物件軌跡以及自車的狀態等時序特徵。接著，輸入至編碼器中進行編碼，此編碼器主要由長短期記憶與多頭自注意力機制組成，分別針對目標物件的時序特徵以及社交關係進行編碼。接著，目的地導向預測器則是透過長短期記憶與多頭自注意力機制先行預測未來軌跡，並分別向前回饋給編碼器以輔助特徵編碼生成；同時向後輸出至後續的條件變分自動編碼器，以用來輔助最終的軌跡預測結果。第三部分為條件變分自動編碼器將未來軌跡做為條件，生成符合條件的未來軌跡多模態(multimodal)分佈。最終透過基於多

頭自注意力機制的解碼器，有效預測出更準確的軌跡路徑。最後本文主要是採用TITAN公開資料庫，以進行本文所發展的演算模型驗證與量化分析。經實驗結果發現，本文所提方法其預測軌跡的平均位移誤差(ADE)能有效改善5%、最終位移誤差(FDE)更能有效改善21%，同時最終交並比(FIOU)也提升9%。

基於單影像之六自由度物體姿態估測

為了解決Auto parts UK的問題，作者洪金利 這樣論述：

Dealing with the object pose estimation from a single RGB image is very challenging since 6 degree-of-freedom (6DoF) parameters have to be predicted without using the spatial depth information. Since direct regression of the pose parameters by using the deep neural network was reportedly poor and t

hen attaching with the refinement module to improve the accuracy causes much time consumption, in this work, we propose several techniques of top-down or bottom-up approaches to predict indirect feature maps instead from which single or multiple object poses can be recovered by using sophisticated p

ost-processing algorithms.Since there are four possible scenarios where single/multiple objects in the same/different classes can appear in the image, the corresponding output feature maps are predicted differently. For a single object scenario, unit-vector fields are predicted. These features are c

omposed of many unit-vectors pointing from pixels within the object mask to the pre-defined 2D object keypoints where their corresponding 3D object keypoints are distributed optimally on the 3D object surface based on the keypoint distances and object surface curvatures. From some pairs of the predi

cted unit-vectors, 2D projected keypoints can be voted and determined, so that PnP algorithm can be applied to estimate the pose. To deal with multiple objects even in the same or different classes, sufficient and informative output feature maps need to be predicted. Different from object keypoints,

6D coordinate maps which form the main features can be considered as a bunch of 3D point clouds for pose parameter calculation when their 2D-3D correspondences are also established. 6D coordinate maps contains two parts: front- and rear-view 3D coordinate maps. 3D coordinate map is actually a 2D ma

p where each pixel records 3D coordinates of a point in the object CAD model which projects to that 2D pixel location. Via 3D/6D coordinate maps, instance 2D-3D correspondences of a large point set can be built and PnP algorithm combined with RANSAC scheme to overcome the outliers or noise can be us

ed to estimate multiple object poses. Even though in this case, 2D object keypoints can no longer be used to estimate multiple poses, they can be defined as single/multiple reference points for identifying all object instance masks even in the presence of heavy occlusion. We are also interested in o

vercoming some problems related to the missing information and symmetry ambiguity encountered when generating the ground truth of 6D coordinate maps.Our studies show that our single pose estimation method using unit-vector fields can achieve an outstanding accuracy if compared to other top-down stat

e-of-the-art methods without including refinement modules. It has a good algorithm to identify the designated object keypoints from which the predicted feature maps are trained with the effective loss functions, but it has a slower inference speed when multiple object poses are taken into considerat

ion. On the other hand, our 6D coordinate maps, combining with the information from two opposite views, are capable of providing more constraints for network optimization and hence helpful for pose estimation accuracy. Our methods using 6D coordinate maps can achieve great performances if compared t

o other multiple object pose estimation methods.