Udacity Sensor Fusion Nanodegree

Data Buffer Optimization

• STL C++ container deque is used to implement the data buffer.

Keypoint Detection

• SHITOMASI, HARRIS, FAST, BRISK, ORB, AKAZE, and SIFT detectors are utilized for features detection and made selectable by setting a string accordingly.

Keypoint Removal

• All keypoints outside of a pre-defined rectangle(preceding vehicle) are removed and only the keypoints within the rectangle are considered for further processing.

Keypoint Descriptors

• BRISK, BRIEF, ORB, FREAK, AKAZE and SIFT descriptors implemented and made selectable by setting a string accordingly.

Descriptor Matching

• FLANN matching as well as k-nearest neighbor selection are implemented. Both methods are selectable using the respective strings in the main function.
• In the Brute force matching, the metric used for matching keypoints is L2 norm for HOG based descriptors(SIFT) and Hamming distance for binary based descriptors(BRIEF, BRISK, ...etc).

Descriptor Distance Ratio

• The 2-Nearest-Neighbor underwent the ratio test, to decide whether the 1-NN is much better than the 2-NN and can be selected as a match or there are many ambigous similar matches and thus 1-NN can be rejected.

Performance Evaluation 1

Number of keypoints detected on the preceding vehicle:

Detector	avg. P.V.K	avg. P.V.K/ avg. T.D.K (%)	Notes on the keypoints
SHI-TOMASI	118	9	well distributed
HARRIS	25	15	well distributed (zero overlapping)
SIFT	139	10	well distributed
FAST	149	8	some clusters and overlapping
BRISK	276	10	some clusters and overlapping
ORB	116	23	some clusters and overlapping
AKAZE	167	12	scattered in a certain area

P.V.K: Keypoints detected on the preceding vehicle.

T.D.K: Total detected keypoints.

---

Performance Evaluation 2

Average number of matched keypoints:

Detector\Descriptor	BRISK	BRIEF	ORB	FREAK	AKAZE	SIFT
SHI-TOMASI	85	105	101	85	_	103
HARRIS	16	19	18	16	_	18
SIFT	66	78	_	66	_	89
FAST	100	122	120	98	_	116
BRISK	174	189	168	170	_	183
ORB	83	61	85	47	_	85
AKAZE	135	141	132	132	140	141

Performance Evaluation 3

Average detector/descriptor processing time:

Detector\Descriptor	BRISK	BRIEF	ORB	FREAK	AKAZE	SIFT
SHI-TOMASI	14.8	13.5	15.8	32	_	29.5
HARRIS	14.8	14	16.2	32.7	_	29.2
SIFT	76.7	76.2	_	95.3	_	126.9
FAST	2.7	1.7	3.5	23.4	_	16.7
BRISK	64.4	61.1	71.5	85.2	_	85.1
ORB	9.1	8	17.3	30.3	_	32.4
AKAZE	63.4	61	69.1	80.5	12.9	73

As expected the binary based detectors/descriptors required less processing time. Based on the recoreded data, The best three detector/descriptor combinations with respect to the total processing time are FAST|BRIEF, FAST|BRISK, and FAST|ORB.