EnrichmentWeek Binus Computer Vision
0 likes22 views
The document discusses computer vision and artificial intelligence. It provides an overview of computer vision applications in AI like object recognition, face recognition, gesture recognition etc. It then discusses different computer vision techniques like stereo cameras, LiDAR, structured light and time-of-flight. It also compares these techniques and discusses their uses in applications like self-driving cars. The document then covers topics like machine learning algorithms for computer vision like convolutional neural networks and different search algorithms used in AI.
![BREADTH FIRST SEARCH ALGORITHM
Python Implementation
graph = {
'S' : ['A','B','C'],
'A' : ['D','S'],
'B' : ['E','S'],
'C' : ['F','S'],
'D' : ['G','A'],
'E' : ['G','B'],
'F' : ['G','C'],
'G' : ['D','E','F]
}
visited = [] # List to keep track of visited nodes.
queue = [] #Initialize a queue
def bfs(visited, graph, node):
visited.append(node)
queue.append(node)
while queue:
s = queue.pop(0)
print (s, end = " ")
for neighbour in graph[s]:
if neighbour not in visited:
visited.append(neighbour)
queue.append(neighbour)
# Driver Code
bfs(visited, graph, 'S')
Open file 1-1_BFS](https://image.slidesharecdn.com/enrichmentweekbinuscv-230813141832-103cd9b0/85/EnrichmentWeek-Binus-Computer-Vision-27-320.jpg)
![DEPTH FIRST SEARCH ALGORITHM
Python Implementation graph = {
'S' : ['A','B','C'],
'A' : ['D','S'],
'B' : ['E','S'],
'C' : ['F','S'],
'D' : ['G','A'],
'E' : ['G','B'],
'F' : ['G','C'],
'G' : ['D', 'E','F']
}
visited = set() # Set to keep track of visited nodes.
def dfs(visited, graph, node):
if node not in visited:
print (node)
visited.add(node)
for neighbour in graph[node]:
dfs(visited, graph, neighbour)
# Driver Code
dfs(visited, graph, 'S')
Open file 1-2_DFS](https://image.slidesharecdn.com/enrichmentweekbinuscv-230813141832-103cd9b0/85/EnrichmentWeek-Binus-Computer-Vision-29-320.jpg)
![WORKSHOP
A* ALGORITHM (INFORMED)
grid = [[0, 1, 0, 0, 0, 0],
[0, 1, 0, 0, 0, 0],
[0, 1, 0, 0, 0, 0],
[0, 1, 0, 0, 0, 0],
[0, 0, 0, 0, 1, 0]]
heuristic = [[9, 8, 7, 6, 5, 4],
[8, 7, 6, 5, 4, 3],
[7, 6, 5, 4, 3, 2],
[6, 5, 4, 3, 2, 1],
[5, 4, 3, 2, 1, 0]]
init = [0, 0]
goal = [len(grid)-1, len(grid[0])-1]
cost = 1
delta = [[-1, 0 ], # go up
[ 0, -1], # go left
[ 1, 0 ], # go down
[ 0, 1 ]] # go right
delta_name = ['^', '<', 'v', '>']
result = search(grid,init,goal,cost,heuristic)
for el in result:
print (el)
def search(grid,init,goal,cost,heuristic):
# ----------------------------------------
# modify the code below
# ----------------------------------------
closed = [[0 for col in range(len(grid[0]))] for row in range(len(grid))]
closed[init[0]][init[1]] = 1
expand = [[-1 for col in range(len(grid[0]))] for row in range(len(grid))]
action = [[-1 for col in range(len(grid[0]))] for row in range(len(grid))]
x = init[0]
y = init[1]
g = 0
f = g + heuristic[x][y]
open = [[f, g, x, y]]
found = False # flag that is set when search is complete
resign = False # flag set if we can't find expand
count = 0
while not found and not resign:
if len(open) == 0:
resign = True
return "Fail"
else:
open.sort()
open.reverse()
next = open.pop()
f = next[0]
g = next[1]
x = next[2]
y = next[3]
expand[x][y] = count
count += 1
if x == goal[0] and y == goal[1]:
found = True
else:
for i in range(len(delta)):
x2 = x + delta[i][0]
y2 = y + delta[i][1]
if x2 >= 0 and x2 < len(grid) and y2 >=0 and y2 < len(grid[0]):
if closed[x2][y2] == 0 and grid[x2][y2] == 0:
g2 = g + cost
f = g2 + heuristic[x2][y2]
open.append([f, g2, x2, y2])
closed[x2][y2] = 1
return expand
Open file 2-1_A_Asterisk](https://image.slidesharecdn.com/enrichmentweekbinuscv-230813141832-103cd9b0/85/EnrichmentWeek-Binus-Computer-Vision-31-320.jpg)
EnrichmentWeek Binus Computer Vision
- 2. Gia Muhammad Direktur Teknologi, Software Engineering and Artificial Intelligence Specialist Xingular.AI PT. Xingular Inovasi Teknologi https://xingular.ai gia@xingular.ai https://www.linkedin.com/in/giamuhammad/ https://github.com/giamuhammad http://ipb.academia.edu/giamuhammad +62819 0865 2920 Vendor Partnership
- 6. ARTIFICIAL INTELLIGENCE Artificial Intelligence is the science and engineering of making machine have an intelligence (McCarthy, 2007) Artificial General Intelligence Artificial Narrow Intelligence Artificial Super Intelligence Technology Singularity
- 7. ARTIFICIAL INTELLIGENCE Intelligence is about tolerance, it able to face uncertainty Human also can make mistakes, yet progress in learning to avoid the same mistake, so neither in AI
- 8. COMPUTER VISION IN ARTIFICIAL INTELLIGENCE Artificial Intelligence Natural Language Processing Computer/Machine Vision Object Recognition Face Recognition User Retention Measurement Security Access and People Tracking Face and Body Analytics Audience Measurement Gesture Recognition Behavior Recognition Object Detection and Tracking Traffic Analytics Environment/Scene Perception Optical Character Recognition ALPR Automatic Document Scanner Barcode 1D Barcode 2D Barcode Augmented Reality Mixed Reality (XR) Face Filter Image Processing Image Editing Watermarking Image Compression Robotics Speech Recognition
- 9. COMPUTER VISION IN ARTIFICIAL INTELLIGENCE Artificial Intelligence Non- Knowledge Based Knowledge Based
- 10. COMPUTER VISION IN ARTIFICIAL INTELLIGENCE Computer Vision Non- Knowledge Based Image Processing Preprocessing Image Preparation Image Feature Visual Perception Pose Estimation View- Geometry Single View Multi View Stereo Vision Knowledge Based Computer Vision Non- Knowledge Based Knowledge Based Object Recognition Machine Learning Non-Neural Net SVM etc Neural Network Deep Learning CNN RCNN
- 11. INTRODUCTION OF SELF-DRIVING CAR Self-Driving car is a vehicle that is capable of sensing its environment and moving safely with little or no human input (Hu; et al, 2020)
- 12. VEHICLE AUTOMATION (VA) LEVEL
- 13. The Human Monitors the Driving Environment The Automated System Monitors the Driving Environment Source : SAE International and J3016 VEHICLE AUTOMATION (VA) LEVEL Artificial Intelligence is still not able to function properly in chaotic inner-city environments. (SingularityU The Netherlands)
- 14. HARDWARE (SENSOR) IN SELF-DRIVING Stereo Camera LiDAR GPS Radar Ultrasonic Wheel Encoder IMU (Inertial Measurement Unit) or INS (Inertial Navigation System)
- 15. STEREO CAMERA
- 16. DEVICE AND METHOD COMPARISON LiDAR for Self-Driving Stereo Camera Cost Expensive Proper Range, Distance Measurement Extensive Limit, Depend on FOV and Geometry Calculation Precision to reconstruct object More Accurate, Sensitive Less Accurate Method Laser Time-of-Flight, Structured Light, Passive and Active Stereo
- 17. DEVICE AND METHOD COMPARISON Laser Structured Light Time-of-Flight Passive Active Method Triangulation Precision Accurate Accurate Less Accurate, Depend on 3D Reconstruction Method Less Accurate, Bad for Dark Condition Less Accurate, Depend on 3D Reconstruction Method/Algorithm Speed < 1s About 2 seconds < 1s Depend on Computation Time of Disparity Map < 1s Cost Very High High Medium Low. Just need more than 1 RGB Camera but need extra compute chip to run disparity map Medium
- 18. LIDAR. LIGHT DETECTION AND RANGING High Precision for Short and Long Range Object Implementation - Mapping for Remote Sensing - Self-Driving Car - 3D Printing from 3D Reconstruction - Robot Navigation
- 19. STRUCTURED LIGHT High Precision for Short Range Object Implementation - 3D Map Guides Robot in Industrial Factory - 3D Printing from 3D Reconstruction
- 20. TIME OF FLIGHT Low Precision for Short Range Object (LiDAR like, using different laser, usually combined with RGB Camera Sensor) Implementation - 3D Printing from 3D Reconstruction - 3D Object Detection - Self-Driving Car - Robot Navigation - Camera Effect - Game Control to Detect Skeleton (Microsoft Kinect) RGB Camera Infrared Laser Raw 3D Reconstructed AI-based algorithm 3D Reconstruction refinement Final Result (High Precision) 3D Reconstructed Illustration of ToF Stereo Camera
- 21. LIDAR IN SELF-DRIVING CAR
- 22. WHY STEREO VISION AND AI Elon Musk in Autonomy Day for Investors, 2019
- 23. WHY STEREO VISION AND AI Less Precision 3D Reconstructed AI (The ability to handle uncertainty) Accurate 3D Object Recognition with Low Cost Sensor
- 25. NON-KNOWLEDGE BASED AI Search Algorithm Blind Breadth First Search (BFS) Depth First Search Informed A* Search
- 26. BREADTH FIRST SEARCH ALGORITHM Only decide first/init node List neighbour node first Using queue to find shortest path Start(S) = ABCDEFG Start(a) = bcdefgh Start(D) = ?
- 27. BREADTH FIRST SEARCH ALGORITHM Python Implementation graph = { 'S' : ['A','B','C'], 'A' : ['D','S'], 'B' : ['E','S'], 'C' : ['F','S'], 'D' : ['G','A'], 'E' : ['G','B'], 'F' : ['G','C'], 'G' : ['D','E','F] } visited = [] # List to keep track of visited nodes. queue = [] #Initialize a queue def bfs(visited, graph, node): visited.append(node) queue.append(node) while queue: s = queue.pop(0) print (s, end = " ") for neighbour in graph[s]: if neighbour not in visited: visited.append(neighbour) queue.append(neighbour) # Driver Code bfs(visited, graph, 'S') Open file 1-1_BFS
- 28. DEPTH FIRST SEARCH ALGORITHM Only decide first/init node List successor of node first Using stack to find shortest path Start(S) = ADGEBFC Start(C) = ?
- 29. DEPTH FIRST SEARCH ALGORITHM Python Implementation graph = { 'S' : ['A','B','C'], 'A' : ['D','S'], 'B' : ['E','S'], 'C' : ['F','S'], 'D' : ['G','A'], 'E' : ['G','B'], 'F' : ['G','C'], 'G' : ['D', 'E','F'] } visited = set() # Set to keep track of visited nodes. def dfs(visited, graph, node): if node not in visited: print (node) visited.add(node) for neighbour in graph[node]: dfs(visited, graph, neighbour) # Driver Code dfs(visited, graph, 'S') Open file 1-2_DFS
- 30. A* ALGORITHM (INFORMED) Shortness Path Finding Algorithm Decide first/init and finish node Considering minimum cost F = g + h where g is distance current node to and the start node, h is estimated distance from current node to end node
- 31. WORKSHOP A* ALGORITHM (INFORMED) grid = [[0, 1, 0, 0, 0, 0], [0, 1, 0, 0, 0, 0], [0, 1, 0, 0, 0, 0], [0, 1, 0, 0, 0, 0], [0, 0, 0, 0, 1, 0]] heuristic = [[9, 8, 7, 6, 5, 4], [8, 7, 6, 5, 4, 3], [7, 6, 5, 4, 3, 2], [6, 5, 4, 3, 2, 1], [5, 4, 3, 2, 1, 0]] init = [0, 0] goal = [len(grid)-1, len(grid[0])-1] cost = 1 delta = [[-1, 0 ], # go up [ 0, -1], # go left [ 1, 0 ], # go down [ 0, 1 ]] # go right delta_name = ['^', '<', 'v', '>'] result = search(grid,init,goal,cost,heuristic) for el in result: print (el) def search(grid,init,goal,cost,heuristic): # ---------------------------------------- # modify the code below # ---------------------------------------- closed = [[0 for col in range(len(grid[0]))] for row in range(len(grid))] closed[init[0]][init[1]] = 1 expand = [[-1 for col in range(len(grid[0]))] for row in range(len(grid))] action = [[-1 for col in range(len(grid[0]))] for row in range(len(grid))] x = init[0] y = init[1] g = 0 f = g + heuristic[x][y] open = [[f, g, x, y]] found = False # flag that is set when search is complete resign = False # flag set if we can't find expand count = 0 while not found and not resign: if len(open) == 0: resign = True return "Fail" else: open.sort() open.reverse() next = open.pop() f = next[0] g = next[1] x = next[2] y = next[3] expand[x][y] = count count += 1 if x == goal[0] and y == goal[1]: found = True else: for i in range(len(delta)): x2 = x + delta[i][0] y2 = y + delta[i][1] if x2 >= 0 and x2 < len(grid) and y2 >=0 and y2 < len(grid[0]): if closed[x2][y2] == 0 and grid[x2][y2] == 0: g2 = g + cost f = g2 + heuristic[x2][y2] open.append([f, g2, x2, y2]) closed[x2][y2] = 1 return expand Open file 2-1_A_Asterisk
- 32. BFS/DFS AND A* FOR ROBOT NAVIGATION WAREHOUSE CASE STUDY Robot Equipped with Stereo Camera or LiDAR Use BFS/DFS to Scan Environment 3D Map Reconstructed The Robot can be navigated using A* pathfinding and 3D Map as parameter When Robot navigated, Stereo Camera/LiDAR also active to avoid obstacle in current time
- 33. KNOWLEDGE-BASED AI DATA SCIENCE Discovery Domain Problem Data Collection Data Preparation Data Visualization (EDA) Feature Engineering/Extraction Data Cleansing Up/Down Sampling Model Planning Regression Classification Clustering Association Model Building Machine Learning Algorithm Tunning Strategy Model Evaluation Operation Report of Model Communication Communicate the Insight to all Stakeholder Data Science Analytics Math, Discrete Math, Statistics Programming Domain/Business Knowledge
- 34. DATA SCIENCE DISCOVERY Domain Problem Understanding Example Problem Government need to know what happen in the current traffic Example Solution Detect all vehicle type Example Insight Personal car is the most vehicle that affect traffic jam in rush hour Example Problem Government need to reduce traffic jam (Decision converted to Problem) Example Solution Detect all plate number Example Insight Odd and even plate number always balanced when traffic jam happen
- 35. DATA SCIENCE DISCOVERY EDA (Exploratory Data Analysis)
- 36. DATA SCIENCE DATA PREPARATION EDA (Exploratory Data Analysis)
- 37. DATA SCIENCE DATA PREPARATION Feature Engineering/Extraction Why data need to be extracted? - Among variables mostly have similar characteristic/feature - Modelling should process data as knowledge to be learned but high dimensional data would make high computational cost - Raw knowledge would make modelling algorithm hard to find the pattern What should we do? - Find the pattern or feature manually - EDA is one of tools to find the pattern
- 38. DATA SCIENCE DATA PREPARATION Feature Engineering/Extraction Lets we jump for a while to model planning and find the pattern Example Case Study - Fish needs to be classified (model planning) into 2 classes/categories - What feature that possible to distinguishing the fishes?
- 39. DATA SCIENCE DATA PREPARATION Feature Engineering/Extraction
- 40. DATA SCIENCE DATA PREPARATION Feature Engineering/Extraction PCA (Principal Component Analysis) - Statistical method to find the feature with dimensionality reduction - Transform first then reduce Steps 1. Standardization (Data Normalization) 2. Find Covariance 3. Find Eigenvalue and Eigenvector from Covariance matrix (cov(X,Y) = A) 4. Sort and choose the Principal Components 5. Transform (Final = Normalized Data * Eigenvector) Open file 3-1_FeatureExtract_PCA_Iris & 3-2_FeatureExtract_PCA_Cifar10
- 41. DATA SCIENCE DATA PREPARATION Feature Engineering/Extraction with Image Processing Edge Detection Line Detection Point Detection Convolution Image Segmentation with Advanced Method Image Segmentation with Basic Method
- 42. DATA SCIENCE DATA PREPARATION Feature Engineering/Extraction with Image Processing Edge Detection Open file 4-1_EdgeDetection | Please use web camera
- 43. DATA SCIENCE DATA PREPARATION Feature Engineering/Extraction with Image Processing Line Detection Open file 4-2_LineDetection | Please use web camera
- 44. DATA SCIENCE DATA PREPARATION Feature Engineering/Extraction with Image Processing Line Detection Open file 4-3_PointDetection | Please use web camera
- 45. DATA SCIENCE DATA PREPARATION Feature Engineering/Extraction with Image Processing Morphology Dilation Erotion
- 46. DATA SCIENCE DATA PREPARATION Feature Engineering/Extraction with Image Processing Image Convolution
- 47. DATA SCIENCE MODEL PLANNING Prediction Type Regression Classification Clustering Association
- 48. DATA SCIENCE MODEL BUILDING How to make a model? Data (x) f(x) = x^2 xi = (3, 5, 2, 1) yi = (9, 25, 4, 1) Known Formula E = mc^2 m = mass c = speed of light Energy Known Formula f(x) = ? xi = (1, 3, 5, 7) yi = (1,3,5,7) Unknown Formula machine learning have an objective to find the formula Classified Target (y) f(x) Blackbox Formula (model)
- 49. DATA SCIENCE MODEL BUILDING How to make a model? Learning Process, Find Minimum Error with iterating method (comparing target data with the current model prediction)
- 50. DATA SCIENCE MODEL BUILDING Model Characteristic in Real-World Problem - No Exact Formula - Always have an error - We never know what formula in math equation Real World Problem
- 51. DATA SCIENCE MODEL BUILDING Minimum Error Trap
- 52. DATA SCIENCE MODEL BUILDING Over-fitting and Under-fitting Generalization is the key We need consider trade-off bias and error Model with the 100% accuracy need to be questioned. (it must be over-fitting) Sometime we should to stop learning process in local minima not global minima to avoid over-fitting
- 53. DATA SCIENCE MODEL BUILDING A sample formula for learning process (Formula to find formula)
- 54. DATA SCIENCE MODEL BUILDING Data Splitting Split Data into training set and test set Using training set to get accuracy Using test set to get validation Data Split Strategy Hold-out (70:30, 80:20) Cross-validation (K-Fold) Evaluation Confusion Matrix Accuracy Precision Recall Specificity F1 Score
- 55. DATA SCIENCE MODEL BUILDING What classifier type to separated the data? Linear? Non-Linear? B A
- 56. DATA SCIENCE MODEL BUILDING How Linear/Logistic Regression find best model Open file 5-1_LinearRegression_Diabetes 5-2_LogisticRegression_Iris 5-3_LogisticRegression_PCA_Cifar10
- 57. DATA SCIENCE MODEL BUILDING Artificial Neural Network (ANN) Neural Network Artificial NN
- 58. DATA SCIENCE MODEL BUILDING Artificial Neural Network (ANN) Perceptron Network Able to create linear classifier
- 59. DATA SCIENCE MODEL BUILDING Artificial Neural Network (ANN) Multilayer Perceptron Network Able to create non-linear classifier Easy to over-fitting Create bias neuron to make model not overfit
- 60. DATA SCIENCE MODEL BUILDING How ANN Learn? Back-propagate error Using difference target with current prediction then consider it to update weight Doing it iteratively until the error descend Open file 5-4_ANN_PCA_Cifar10
- 61. DATA SCIENCE MODEL BUILDING Convolution Neural Network
- 62. DATA SCIENCE MODEL BUILDING Deep Learning Different Activation Function More layer (Deep) Different Backpropagation Algorithm
- 63. DATA SCIENCE MODEL BUILDING Activation Function and Backpropagation algorithm comparation
- 64. DATA SCIENCE MODEL BUILDING Deep Convolution Neural Network Open file 5-4_DCNN_Cifar10
- 65. DATA SCIENCE MODEL BUILDING Deep Convolution Neural Network Standardized DCNN Architecture for Object Recognition ResNet (Residual Network) by AlexNet (Team name) win the ImageNet 2012 competition VGG b Simonyan and Zisserman
- 67. DATA SCIENCE MODEL BUILDING Deep Convolution Neural Network Standarized DCNN Architecture for Object Recognition Inception Architecture by Google
- 68. DATA SCIENCE MODEL BUILDING Deep Convolution Neural Network Standardized DCNN Architecture for Object Recognition
- 69. DATA SCIENCE MODEL BUILDING Regional-Convolution Neural Network
- 70. COMPUTER VISION WITH DEEP LEARNING IMPLEMENTATION 3D (Object) Regional-Convolution Neural Network for Self-Driving Car
- 71. COMPUTER VISION WITH DEEP LEARNING IMPLEMENTATION
- 72. EDGE COMPUTING FOR AIOT