Javelin-throw analysis

Uncovering the objective

The primary objective in the sport of javelin throw is to achieve the maximum distance of flight of javelin. This distance traveled by the javelin is dependent mainly on the angle of release of the javelin and the approach run to gain velocity.

The javelin throw is different from all other overhead throws as it is more of an over-arm whip and flip motion that uses the efforts of the entire body. Terms such as fling, flip, the whip are much more descriptive and result in a more relaxed sequentially efficient delivery in which the arm of the athlete gets involved in the projection only after the major muscles of the legs, hips, and trunk have been utilized.

The best javelin throw of all time was performed by Jan Zelezny, of the Czech Republic in 1966. The result was 98.48m. In the second place, comes Johannes Vetter of Germany in 2017, with a distance of 94.44m. In third place, comes Thomas Rohler, also from Germany, with a throw of 93.90m, in 2017.

The flight distance of the javelin is determined by the release parameters such as the angle of release, the height of release, and velocity of release, as well as by the forces that act on the javelin during its flight. The former is under the control of the thrower, whereas the latter is not.

The objective here is to find the optimal body posture of the athlete at each step during the approach phase for achieving maximum javelin flight. We make use of an open-source framework, called Mediapipe, introduced by Google to build a machine learning model that analyzes the body posture of the athlete at each step in the approach phase and the release phase.

Methodology

The body posture of top javelin throw athletes was measured by building a machine learning model that estimates the body posture of an athlete in each frame of videos collected from the internet. These athletes include:

• Neeraj Chopra
  
• Johannes Vetter
  
• Thomas Rohler
  
• Andreas Hofmann
  
• Magnus Kirt
  

Since the information such as the number of steps taken in the approach phase varied from one video to another, all the videos are trimmed such that each of them consists of the last six steps, which is generally referred to as the crossover phase. After undergoing trimming, we remain with videos that have 90 frames each. According to studies conducted at Concordia University, NE, on analyzing the biomechanics behind javelin throw, seven main muscle groups play a major role in determining the velocity gained by the athlete in the approach phase. These muscle groups are as follows:

• Gastrocnemius muscle
  
• Soleus muscle
  
• Quadriceps
  
• Hamstrings
  
• Rectus Abdominis
  
• Triceps
  
• Rhomboids
  

To measure the position of these muscle groups, we calculate 8 different angles from the mediapipe Pose framework. The landmarks of these angles are as shown in the image below:

The pose-estimation model is then implemented, which iterates through all the videos to calculate the eight different angles of each frame of the video, and these angles collected are stored in separate CSV(comma separated values) files. After this process, the angles obtained from each video are analyzed to find the position of the athlete when each of the 6 steps in the crossover phase is placed and the position of the athlete during the release phase.

These angles at each step in the approach phase are stored in 7 separate CSV file which is used for further analysis. Each of the final CSV files that are used to derive the optimal position contains the following features:

• Left hip angle
• Right hip angle
• Left knee angle
• Right knee angle
• Left trunk angle
• Right trunk angle
• Elbow angle
• Shoulder angle
• Height of athlete
• Weight of athlete
• Distance traveled by javelin

After this process, regression models were built for all the CSV files of the 7 steps to find the optimal body position to achieve a different range of distances.

Results

The Regression models that were built to find the optimal body position at each of the 7 steps are then used to predict the optimal position of the 7 muscle groups to achieve a different range of distances. The table below shows the different body positions to achieve a range of distances:

At step 1:

Distance (m)	Height (m)	Weight (kg)	Right Hip	Left Hip	Right knee	Left knee	Left trunk	Right trunk	Elbow	Shoulder
85-87	1.8	86	100	90	151	100	163	166	155	36
	1.85	91	102	102	175	121.08	154	146	167	117
	1.9	95	110	105	140	90	173	176	155	40
88-90	1.8	86	105	110	175	122	144	116	162	111
	1.85	91	102	105	173	121.7	134	104	152	122
	1.9	95	106	111	167	141	143	107	154	112
91-93	1.8	86	103	107	176	144	153	107	127	101
	1.85	91	106	112	166	143	154	104	121	86
	1.9	95	120	90	171	147	156	165	134	82

At step 2:

Distance	Height	Weight	Right Hip	Left Hip	Right knee	Left knee	Left trunk	Right trunk	Elbow	Shoulder
82-84	1.8	86	71	97	138	150	160	160	166	70
	1.85	91	80	67	138	144	152	147	174	68
	1.9	95	84	63	131	139	152	157	176	65
85-87	1.8	86	87	74	145	141	163	126	169	102
	1.85	91	92	104	153	147	167	137	167	99
	1.9	95	90	97	127	132	162	138	169	101
88-90	1.8	86	116	94	122	165	144	131	143	81
	1.85	91	96	104	142	133	166	127	131	79
	1.9	95	70	99	162	154	146	132	143	76
91-93	1.8	86	118	107	144	167	137	143	147	84
	1.9	95	90	97	157	132	172	148	145	101

At step 3:

Distance	Height	Weight	Right Hip	Left Hip	Right knee	Left knee	Left trunk	Right trunk	Elbow	Shoulder
82-84	1.8	86	84	103	138	161	142	159	161	79
	1.85	91	86	99	132	153	147	161	160	69
	1.9	95	89	97	127	149	154	163	162	67
85-87	1.8	86	96	105	157	142	173	154	157	74
	1.85	91	104	98	149	147	169	167	159	86
	1.9	95	102	101	152	151	170	159	155	92
88-90	1.8	86	99	104	157	131	145	162	152	102
	1.85	91	101	102	149	137	160	169	149	87
	1.9	95	110	103	142	139	163	157	152	96
91-93	1.8	86	114	106	157	138	125	177	152	57
	1.9	95	116	112	152	134	132	159	153	61

At step 4:

Distance	Height	Weight	Right Hip	Left Hip	Right knee	Left knee	Left trunk	Right trunk	Elbow	Shoulder
82-84	1.8	86	94	92	138	139	150	159	174	52
	1.85	91	104	122	131	145	152	149	174	61
	1.9	95	99	110	134	143	149	153	173	72
85-87	1.8	86	97	139	141	148	132	154	169	102
	1.85	91	101	123	147	142	150	152	172	94
	1.9	95	92	132	151	138	152	147	171	104
88-90	1.8	86	97	111	132	147	153	147	163	79
	1.85	91	100	141	166	143	118	168	162	72
	1.9	95	110	121	149	136	144	151	159	92
91-93	1.8	86	104	121	139	151	169	148	144	101
	1.9	95	103	114	161	149	175	154	159	104

At step 5:

Distance	Height	Weight	Right Hip	Left Hip	Right knee	Left knee	Left trunk	Right trunk	Elbow	Shoulder
82-84	1.8	86	115	107	163	129	164	166	161	72
	1.85	91	111	114	147	149	159	156	172	69
	1.9	95	99	121	161	139	163	163	167	55
85-87	1.8	86	97	114	164	145	161	150	161	75
	1.85	91	92	106	156	152	159	157	159	72
	1.9	95	104	99	159	147	152	157	162	77
88-90	1.8	86	113	100	162	157	149	149	151	76
	1.85	91	99	105	146	161	149	152	153	69
	1.9	95	104	103	151	159	150	152	151	71
91-93	1.8	86	99	95	151	167	148	145	141	66
	1.9	95	97	99	147	165	150	151	146	72

At step 6 (final step):

Distance	Height	Weight	Right Hip	Left Hip	Right knee	Left knee	Left trunk	Right trunk	Elbow	Shoulder
82-84	1.8	86	91	121	163	160	149	169	139	82
	1.85	91	84	119	163	165	154	172	132	110
	1.9	95	80	152	164	152	136	177	122	143
85-87	1.8	86	76	113	164	170	162	164	136	111
	1.85	91	73	104	161	168	162	171	140	96
	1.9	95	69	132	159	173	153	167	134	76
88-90	1.8	86	76	97	169	171	166	164	121	109
	1.85	91	74	93	165	167	161	164	132	156
	1.9	95	77	121	162	176	169	174	142	121
91-93	1.8	86	77	92	169	172	168	164	152	72
	1.9	95	72	86	167	169	154	167	143	99

During release:

Distance	Height	Weight	Right Hip	Left Hip	Right knee	Left knee	Left trunk	Right trunk	Elbow	Shoulder
82-84	1.8	86	75	127	142	176	133	159	165	164
	1.85	91	77	125	156	167	152	164	166	172
	1.9	95	79	112	144	172	163	168	160	161
85-87	1.8	86	67	142	167	162	154	166	164	169
	1.85	91	76	119	149	170	164	170	165	164
	1.9	95	74	121	139	163	166	168	159	170
88-90	1.8	86	66	100	164	174	162	169	168	152
91-93	1.8	86	76	131	142	173	169	163	162	159
96	1.8	86	104	125	132	172	177	164	165	177

Conclusion

The use of pose estimation machine learning model has been demonstrated to simulate the flight distance of javelins. This approach can help in training the athletes to attain better performance by correcting the basic body posture during the approach phase. The body pose can be adjusted by referencing the above charts at each step to attain better distances.

The results obtained after training the machine learning models seem to be encouragingly accurate, taking into consideration the amount of training sets.